diff --git a/.gitignore b/.gitignore
index 66c046a0b5..88de384705 100644
--- a/.gitignore
+++ b/.gitignore
@@ -217,4 +217,5 @@ inference_testing
 *.rrd
 
 openspec*/
-opsx/
\ No newline at end of file
+opsx/
+checkpoints/
diff --git a/docker/dockerfiles/Dockerfile.onnx.gpu b/docker/dockerfiles/Dockerfile.onnx.gpu
index 33969fbb8a..58f925da47 100644
--- a/docker/dockerfiles/Dockerfile.onnx.gpu
+++ b/docker/dockerfiles/Dockerfile.onnx.gpu
@@ -108,6 +108,10 @@ RUN ln -s /usr/bin/python3 /usr/bin/python
 RUN /bin/make create_wheels_for_gpu_notebook
 RUN pip3 install --no-cache-dir dist/inference_cli*.whl dist/inference_core*.whl dist/inference_gpu*.whl dist/inference_sdk*.whl "setuptools<=75.5.0"
 
+# Instal Cosmos 3 preview (requires override of transformers, since 21.07.2026 it was not released yet).
+COPY requirements/requirements.cosmos.txt .
+RUN python3 -m pip install -r requirements.cosmos.txt
+
 WORKDIR /notebooks
 COPY examples/notebooks .
 
diff --git a/inference/core/env.py b/inference/core/env.py
index 7d8b1ed404..445eb12af8 100644
--- a/inference/core/env.py
+++ b/inference/core/env.py
@@ -259,6 +259,8 @@
 
 LMM_ENABLED = str2bool(os.getenv("LMM_ENABLED", False))
 
+COSMOS3_ENABLED = str2bool(os.getenv("COSMOS3_ENABLED", True))
+
 QWEN_2_5_ENABLED = str2bool(os.getenv("QWEN_2_5_ENABLED", True))
 
 QWEN_3_ENABLED = str2bool(os.getenv("QWEN_3_ENABLED", True))
diff --git a/inference/core/workflows/core_steps/loader.py b/inference/core/workflows/core_steps/loader.py
index f6d6df010c..9675f575c1 100644
--- a/inference/core/workflows/core_steps/loader.py
+++ b/inference/core/workflows/core_steps/loader.py
@@ -238,6 +238,9 @@
 from inference.core.workflows.core_steps.models.foundation.cog_vlm.v1 import (
     CogVLMBlockV1,
 )
+from inference.core.workflows.core_steps.models.foundation.cosmos3.v1 import (
+    Cosmos3EdgeBlockV1,
+)
 from inference.core.workflows.core_steps.models.foundation.depth_estimation.v1 import (
     DepthEstimationBlockV1,
 )
@@ -1009,6 +1012,7 @@ def load_blocks() -> List[Type[WorkflowBlock]]:
         GoogleGemmaBlockV1,
         GoogleGemmaBlockV2,
         ImageSlicerBlockV2,
+        Cosmos3EdgeBlockV1,
         Qwen25VLBlockV1,
         Qwen3VLBlockV1,
         Qwen35VLBlockV1,
diff --git a/inference/core/workflows/core_steps/models/foundation/cosmos3/__init__.py b/inference/core/workflows/core_steps/models/foundation/cosmos3/__init__.py
new file mode 100644
index 0000000000..e69de29bb2
diff --git a/inference/core/workflows/core_steps/models/foundation/cosmos3/v1.py b/inference/core/workflows/core_steps/models/foundation/cosmos3/v1.py
new file mode 100644
index 0000000000..f46d939ce1
--- /dev/null
+++ b/inference/core/workflows/core_steps/models/foundation/cosmos3/v1.py
@@ -0,0 +1,254 @@
+from typing import List, Literal, Optional, Type, Union
+
+from pydantic import ConfigDict, Field
+
+from inference.core.entities.requests.inference import LMMInferenceRequest
+from inference.core.env import (
+    COSMOS3_ENABLED,
+    HOSTED_CORE_MODEL_URL,
+    LOCAL_INFERENCE_API_URL,
+    WORKFLOWS_REMOTE_API_TARGET,
+)
+from inference.core.managers.base import ModelManager
+from inference.core.workflows.core_steps.common.entities import StepExecutionMode
+from inference.core.workflows.execution_engine.entities.base import (
+    Batch,
+    OutputDefinition,
+    WorkflowImageData,
+)
+from inference.core.workflows.execution_engine.entities.types import (
+    IMAGE_KIND,
+    LANGUAGE_MODEL_OUTPUT_KIND,
+    ROBOFLOW_MODEL_ID_KIND,
+    STRING_KIND,
+    ImageInputField,
+    Selector,
+)
+from inference.core.workflows.prototypes.block import (
+    BlockResult,
+    Runtime,
+    RuntimeRestriction,
+    Severity,
+    WorkflowBlock,
+    WorkflowBlockManifest,
+)
+from inference_sdk import InferenceHTTPClient
+
+DEFAULT_PROMPT = "Describe what's in this image."
+DEFAULT_SYSTEM_PROMPT = (
+    "You are Cosmos, a helpful assistant that understands physical scenes "
+    "and answers questions about images and videos."
+)
+
+
+class BlockManifest(WorkflowBlockManifest):
+    model_config = ConfigDict(
+        json_schema_extra={
+            "name": "Cosmos 3",
+            "version": "v1",
+            "short_description": "Reason about physical scenes with NVIDIA Cosmos 3.",
+            "long_description": (
+                "This workflow block runs the NVIDIA Cosmos 3 reasoner — a world model "
+                "tuned for physical scene understanding—on an image with an optional text "
+                "prompt, and returns a text answer. The model is well suited to questions "
+                "about spatial relations, safety, and what will happen next in a scene."
+            ),
+            "license": "OpenMDW-1.1",
+            "block_type": "model",
+            "search_keywords": [
+                "Cosmos",
+                "cosmos3",
+                "NVIDIA",
+                "world model",
+                "vision language model",
+                "VLM",
+                "robotics",
+            ],
+            "is_vlm_block": True,
+            "ui_manifest": {
+                "section": "model",
+                "icon": "fal fa-atom",
+                "blockPriority": 5.5,
+            },
+        },
+        protected_namespaces=(),
+    )
+    type: Literal["roboflow_core/cosmos3_edge@v1"]
+
+    images: Selector(kind=[IMAGE_KIND]) = ImageInputField
+    prompt: Optional[Union[Selector(kind=[STRING_KIND]), str]] = Field(
+        default=None,
+        description="Optional text prompt to pass to Cosmos 3 Edge. Otherwise a default "
+        "scene-description prompt is used, which may affect the desired model behavior.",
+        examples=["Is the walkway free of obstacles?", "$inputs.prompt"],
+    )
+    model_version: Union[
+        Selector(kind=[ROBOFLOW_MODEL_ID_KIND]), Literal["nvidia/cosmos-3-edge"]
+    ] = Field(
+        default="nvidia/cosmos-3-edge",
+        description="The Cosmos 3 Edge model to be used for inference.",
+        examples=["nvidia/cosmos-3-edge"],
+    )
+    system_prompt: Optional[Union[Selector(kind=[STRING_KIND]), str]] = Field(
+        default=None,
+        description="Optional system prompt to provide additional context to Cosmos 3 Edge.",
+        examples=["You are a safety inspector.", "$inputs.system_prompt"],
+    )
+
+    @classmethod
+    def describe_outputs(cls) -> List[OutputDefinition]:
+        return [
+            OutputDefinition(
+                name="output",
+                kind=[STRING_KIND, LANGUAGE_MODEL_OUTPUT_KIND],
+                description="The model's text answer for each input image.",
+            ),
+        ]
+
+    @classmethod
+    def get_parameters_accepting_batches(cls) -> List[str]:
+        return ["images"]
+
+    @classmethod
+    def get_execution_engine_compatibility(cls) -> Optional[str]:
+        return ">=1.3.0,<2.0.0"
+
+    @classmethod
+    def get_restrictions(cls) -> List[RuntimeRestriction]:
+        restrictions = [
+            RuntimeRestriction(
+                severity=Severity.HARD,
+                note="Requires a GPU; run_locally() loads a model that needs CUDA.",
+                applies_to_runtimes=[Runtime.SELF_HOSTED_CPU],
+                applies_to_step_execution_modes=[StepExecutionMode.LOCAL],
+            ),
+        ]
+        if not COSMOS3_ENABLED:
+            restrictions.append(
+                RuntimeRestriction(
+                    severity=Severity.HARD,
+                    note=(
+                        "COSMOS3_ENABLED=False: the Cosmos 3 Edge endpoint is not "
+                        "registered, so run_remotely() returns 404."
+                    ),
+                    applies_to_runtimes=[Runtime.HOSTED_SERVERLESS],
+                    applies_to_step_execution_modes=[StepExecutionMode.REMOTE],
+                )
+            )
+        return restrictions
+
+    @classmethod
+    def get_supported_model_variants(cls) -> Optional[List[str]]:
+        """Return list of model_id variants that can satisfy this block."""
+        return ["nvidia/cosmos-3-edge"]
+
+
+class Cosmos3EdgeBlockV1(WorkflowBlock):
+    def __init__(
+        self,
+        model_manager: ModelManager,
+        api_key: Optional[str],
+        step_execution_mode: StepExecutionMode,
+    ):
+        self._model_manager = model_manager
+        self._api_key = api_key
+        self._step_execution_mode = step_execution_mode
+
+    @classmethod
+    def get_init_parameters(cls) -> List[str]:
+        return ["model_manager", "api_key", "step_execution_mode"]
+
+    @classmethod
+    def get_manifest(cls) -> Type[WorkflowBlockManifest]:
+        return BlockManifest
+
+    def run(
+        self,
+        images: Batch[WorkflowImageData],
+        model_version: str,
+        prompt: Optional[str],
+        system_prompt: Optional[str],
+    ) -> BlockResult:
+        if self._step_execution_mode == StepExecutionMode.LOCAL:
+            return self.run_locally(
+                images=images,
+                model_version=model_version,
+                prompt=prompt,
+                system_prompt=system_prompt,
+            )
+        elif self._step_execution_mode == StepExecutionMode.REMOTE:
+            return self.run_remotely(
+                images=images,
+                model_version=model_version,
+                prompt=prompt,
+                system_prompt=system_prompt,
+            )
+        else:
+            raise ValueError(
+                f"Unknown step execution mode: {self._step_execution_mode}"
+            )
+
+    def run_remotely(
+        self,
+        images: Batch[WorkflowImageData],
+        model_version: str,
+        prompt: Optional[str],
+        system_prompt: Optional[str],
+    ) -> BlockResult:
+        api_url = (
+            LOCAL_INFERENCE_API_URL
+            if WORKFLOWS_REMOTE_API_TARGET != "hosted"
+            else HOSTED_CORE_MODEL_URL
+        )
+        client = InferenceHTTPClient(
+            api_url=api_url,
+            api_key=self._api_key,
+        )
+        if WORKFLOWS_REMOTE_API_TARGET == "hosted":
+            client.select_api_v0()
+
+        combined_prompt = _combine_prompt(prompt=prompt, system_prompt=system_prompt)
+        predictions = []
+        for image in images:
+            result = client.infer_lmm(
+                inference_input=image.base64_image,
+                model_id=model_version,
+                prompt=combined_prompt,
+                model_id_in_path=True,
+            )
+            predictions.append({"output": result.get("response", result)})
+        return predictions
+
+    def run_locally(
+        self,
+        images: Batch[WorkflowImageData],
+        model_version: str,
+        prompt: Optional[str],
+        system_prompt: Optional[str],
+    ) -> BlockResult:
+        inference_images = [
+            i.to_inference_format(numpy_preferred=False) for i in images
+        ]
+        combined_prompt = _combine_prompt(prompt=prompt, system_prompt=system_prompt)
+        self._model_manager.add_model(model_id=model_version, api_key=self._api_key)
+
+        predictions = []
+        for image in inference_images:
+            request = LMMInferenceRequest(
+                api_key=self._api_key,
+                model_id=model_version,
+                image=image,
+                source="workflow-execution",
+                prompt=combined_prompt,
+            )
+            prediction = self._model_manager.infer_from_request_sync(
+                model_id=model_version, request=request
+            )
+            predictions.append({"output": prediction.response})
+        return predictions
+
+
+def _combine_prompt(prompt: Optional[str], system_prompt: Optional[str]) -> str:
+    prompt = prompt or DEFAULT_PROMPT
+    system_prompt = system_prompt or DEFAULT_SYSTEM_PROMPT
+    return prompt + "<system_prompt>" + system_prompt
diff --git a/inference/models/README.md b/inference/models/README.md
index f1fff8821f..aff5662266 100644
--- a/inference/models/README.md
+++ b/inference/models/README.md
@@ -3,6 +3,7 @@ The models supported by Roboflow Inference have their own licenses. View the lic
 | model                                 |                                                                           license                                                                            | commercial license available |
 |:--------------------------------------|:------------------------------------------------------------------------------------------------------------------------------------------------------------:|:----------------------------:|
 | `inference/models/clip`               |                                                   [MIT](https://github.com/openai/CLIP/blob/main/LICENSE)                                                    |              👍              |
+| `inference/models/cosmos3`            |                                                    [OpenMDW-1.1](https://openmdw.ai/license/1-1/)                                                    |              👍              |
 | `inference/models/doctr`              |                                               [Apache 2.0](https://github.com/mindee/doctr/blob/main/LICENSE)                                                |              👍              |
 | `inference/models/depth_anything_v2`  |                                      [Apache 2.0](https://github.com/DepthAnything/Depth-Anything-V2/blob/main/LICENSE)                                      |              👍              |
 | `inference/models/depth_anything_v3`  |                                      [Apache 2.0](https://github.com/ByteDance-Seed/depth-anything-3/blob/main/LICENSE)                                      |              👍              |
diff --git a/inference/models/cosmos3/LICENSE.txt b/inference/models/cosmos3/LICENSE.txt
new file mode 100644
index 0000000000..52f32cc9c1
--- /dev/null
+++ b/inference/models/cosmos3/LICENSE.txt
@@ -0,0 +1,55 @@
+NVIDIA Cosmos 3 Edge is released under the OpenMDW License Agreement,
+version 1.1 (OpenMDW-1.1): https://openmdw.ai/license/1-1/
+
+----------------------------------------------------------------------
+
+OpenMDW License Agreement, version 1.1 (OpenMDW-1.1)
+
+By exercising rights granted to you under this agreement, you accept and
+agree to its terms.
+
+As used in this agreement, "Model Materials" means the materials provided
+to you under this agreement, consisting of: (1) one or more machine
+learning models (including architecture and parameters); and (2) all
+related artifacts (including associated data, documentation and software)
+that are provided to you hereunder.
+
+Subject to your compliance with this agreement, permission is hereby
+granted, free of charge, to deal in the Model Materials without
+restriction, including under all copyright, patent, database, and trade
+secret rights included or embodied therein.
+
+If you distribute any portion of the Model Materials, you shall retain in
+your distribution (1) a copy of this agreement, and (2) all copyright
+notices and other notices of origin included in the Model Materials that
+are applicable to your distribution.
+
+If you file, maintain, or voluntarily participate in a lawsuit against
+any person or entity asserting that the Model Materials directly or
+indirectly infringe any patent or copyright, then all rights and grants
+made to you hereunder are terminated, unless that lawsuit was in response
+to a corresponding lawsuit first brought against you.
+
+This agreement does not impose any restrictions or obligations with
+respect to any use, modification, or sharing of any outputs generated by
+using the Model Materials.
+
+THE MODEL MATERIALS ARE PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, TITLE, NONINFRINGEMENT,
+ACCURACY, OR THE ABSENCE OF LATENT OR OTHER DEFECTS OR ERRORS, WHETHER OR
+NOT DISCOVERABLE, ALL TO THE GREATEST EXTENT PERMISSIBLE UNDER APPLICABLE
+LAW.
+
+YOU ARE SOLELY RESPONSIBLE FOR (1) CLEARING RIGHTS OF OTHER PERSONS THAT
+MAY APPLY TO THE MODEL MATERIALS OR ANY USE THEREOF, INCLUDING WITHOUT
+LIMITATION ANY PERSON'S COPYRIGHTS OR OTHER RIGHTS INCLUDED OR EMBODIED IN
+THE MODEL MATERIALS; (2) OBTAINING ANY NECESSARY CONSENTS, PERMISSIONS OR
+OTHER RIGHTS REQUIRED FOR ANY USE OF THE MODEL MATERIALS; OR (3)
+PERFORMING ANY DUE DILIGENCE OR UNDERTAKING ANY OTHER INVESTIGATIONS INTO
+THE MODEL MATERIALS OR ANYTHING INCORPORATED OR EMBODIED THEREIN.
+
+IN NO EVENT SHALL THE PROVIDERS OF THE MODEL MATERIALS BE LIABLE FOR ANY
+CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT
+OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE MODEL
+MATERIALS, THE USE THEREOF OR OTHER DEALINGS THEREIN.
diff --git a/inference/models/cosmos3/__init__.py b/inference/models/cosmos3/__init__.py
new file mode 100644
index 0000000000..e69de29bb2
diff --git a/inference/models/cosmos3/cosmos3_reasoner_inference_models.py b/inference/models/cosmos3/cosmos3_reasoner_inference_models.py
new file mode 100644
index 0000000000..331b2686b6
--- /dev/null
+++ b/inference/models/cosmos3/cosmos3_reasoner_inference_models.py
@@ -0,0 +1,102 @@
+from typing import Any, List
+
+import torch
+
+from inference.core.entities.responses import (
+    InferenceResponseImage,
+    LMMInferenceResponse,
+)
+from inference.core.env import (
+    ALLOW_INFERENCE_MODELS_DIRECTLY_ACCESS_LOCAL_PACKAGES,
+    ALLOW_INFERENCE_MODELS_UNTRUSTED_PACKAGES,
+    API_KEY,
+    DISABLED_INFERENCE_MODELS_BACKENDS,
+    VALID_INFERENCE_MODELS_BACKENDS,
+)
+from inference.core.models.base import Model
+from inference.core.models.types import PreprocessReturnMetadata
+from inference.core.roboflow_api import get_extra_weights_provider_headers
+from inference.core.utils.image_utils import load_image_bgr
+from inference_models import AutoModel, PreProcessingOverrides
+from inference_models.models.cosmos3.cosmos3_reasoner_hf import Cosmos3EdgeReasoner
+
+
+class InferenceModelsCosmos3ReasonerAdapter(Model):
+    def __init__(self, model_id: str, api_key: str = None, **kwargs):
+        super().__init__()
+
+        self.metrics = {"num_inferences": 0, "avg_inference_time": 0.0}
+
+        self.api_key = api_key if api_key else API_KEY
+
+        self.task_type = "lmm"
+
+        extra_weights_provider_headers = get_extra_weights_provider_headers(
+            countinference=kwargs.get("countinference"),
+            service_secret=kwargs.get("service_secret"),
+        )
+        backend = list(
+            VALID_INFERENCE_MODELS_BACKENDS.difference(
+                DISABLED_INFERENCE_MODELS_BACKENDS
+            )
+        )
+        self._model: Cosmos3EdgeReasoner = AutoModel.from_pretrained(
+            model_id_or_path=model_id,
+            api_key=self.api_key,
+            allow_untrusted_packages=ALLOW_INFERENCE_MODELS_UNTRUSTED_PACKAGES,
+            allow_direct_local_storage_loading=ALLOW_INFERENCE_MODELS_DIRECTLY_ACCESS_LOCAL_PACKAGES,
+            weights_provider_extra_headers=extra_weights_provider_headers,
+            backend=backend,
+            **kwargs,
+        )
+
+    def map_inference_kwargs(self, kwargs: dict) -> dict:
+        pre_processing_overrides = PreProcessingOverrides(
+            disable_contrast_enhancement=kwargs.get("disable_preproc_contrast", False),
+            disable_grayscale=kwargs.get("disable_preproc_grayscale", False),
+            disable_static_crop=kwargs.get("disable_preproc_static_crop", False),
+        )
+        kwargs["pre_processing_overrides"] = pre_processing_overrides
+        return kwargs
+
+    def preprocess(self, image: Any, prompt: str = "", **kwargs):
+        is_batch = isinstance(image, list)
+        if is_batch:
+            raise ValueError("This model does not support batched-inference.")
+        np_image = load_image_bgr(
+            image,
+            disable_preproc_auto_orient=kwargs.get(
+                "disable_preproc_auto_orient", False
+            ),
+        )
+        input_shape = PreprocessReturnMetadata({"image_dims": np_image.shape[:2][::-1]})
+        mapped_kwargs = self.map_inference_kwargs(kwargs)
+        return (
+            self._model.pre_process_generation(np_image, prompt, **mapped_kwargs),
+            input_shape,
+        )
+
+    def predict(self, inputs, **kwargs) -> torch.Tensor:
+        mapped_kwargs = self.map_inference_kwargs(kwargs)
+        return self._model.generate(inputs, **mapped_kwargs)
+
+    def postprocess(
+        self,
+        predictions: torch.Tensor,
+        preprocess_return_metadata: PreprocessReturnMetadata,
+        **kwargs,
+    ) -> List[LMMInferenceResponse]:
+        mapped_kwargs = self.map_inference_kwargs(kwargs)
+        result = self._model.post_process_generation(predictions, **mapped_kwargs)[0]
+        return [
+            LMMInferenceResponse(
+                response=result,
+                image=InferenceResponseImage(
+                    width=preprocess_return_metadata["image_dims"][0],
+                    height=preprocess_return_metadata["image_dims"][1],
+                ),
+            )
+        ]
+
+    def clear_cache(self, delete_from_disk: bool = True) -> None:
+        pass
diff --git a/inference/models/utils.py b/inference/models/utils.py
index 16436a5425..04969e3bec 100644
--- a/inference/models/utils.py
+++ b/inference/models/utils.py
@@ -15,6 +15,7 @@
     CORE_MODEL_SAM_ENABLED,
     CORE_MODEL_TROCR_ENABLED,
     CORE_MODEL_YOLO_WORLD_ENABLED,
+    COSMOS3_ENABLED,
     DEPTH_ESTIMATION_ENABLED,
     FLORENCE2_ENABLED,
     GLM_OCR_ENABLED,
@@ -494,6 +495,31 @@
         category=ModelDependencyMissing,
     )
 
+try:
+    # Cosmos 3 Edge has no legacy implementation — it is served exclusively
+    # through the inference_models bridge adapter.
+    if COSMOS3_ENABLED and USE_INFERENCE_MODELS:
+        from inference.models.cosmos3.cosmos3_reasoner_inference_models import (
+            InferenceModelsCosmos3ReasonerAdapter,
+        )
+
+        cosmos3_models = {
+            (
+                "text-image-pairs",
+                "cosmos-3-edge",
+            ): InferenceModelsCosmos3ReasonerAdapter,
+            ("vlm", "cosmos-3-edge"): InferenceModelsCosmos3ReasonerAdapter,
+        }
+        ROBOFLOW_MODEL_TYPES.update(cosmos3_models)
+except:
+    warnings.warn(
+        "Your `inference` configuration does not support the Cosmos 3 model. "
+        "Since inference 1.3.6 was shipped when downstream model dependencies were not released yet, "
+        "we have enabled the model in selected builds only. Installation guide will be provided "
+        "in following releases. To suppress this warning, set COSMOS3_ENABLED to False.",
+        category=ModelDependencyMissing,
+    )
+
 
 try:
     if CORE_MODEL_SAM_ENABLED:
diff --git a/inference_models/development/cosmos3/pull_weights.py b/inference_models/development/cosmos3/pull_weights.py
new file mode 100644
index 0000000000..e84bcf9412
--- /dev/null
+++ b/inference_models/development/cosmos3/pull_weights.py
@@ -0,0 +1,215 @@
+"""Pull NVIDIA Cosmos 3 Edge weights and materialize inference_models package layouts.
+
+The HF repo (nvidia/Cosmos3-Edge) is dual-format: a transformers-style reasoner at
+the repo root and a diffusers-style generator (model_index.json + transformer/ +
+vae/ + scheduler/). The two towers share weight files (the reasoner's
+model.safetensors.index.json maps into transformer/*.safetensors and
+vision_encoder/model.safetensors), so this script downloads ONE snapshot and
+materializes each tower's package directory from it with hardlinks (falling back
+to copies across filesystems).
+
+Package layouts produced (what `from_pretrained` expects):
+
+    <output-dir>/cosmos-3-edge/          -> Cosmos3EdgeReasoner.from_pretrained(...)
+    <output-dir>/cosmos-3-edge-world/    -> Cosmos3EdgeWorldModel.from_pretrained(...)
+
+The world package additionally needs a `cosmos3_generator_runtime.py` at its root
+(the self-contained runtime module the loader imports). Pass --runtime-module to
+inject one; the reference implementation lives next to this script
+(reference_generator_runtime.py).
+
+To mirror a package into GCS (same layout, no wrapping directory), pass
+--gcs-dest; each package uploads to <gcs-dest>/<package-name>/:
+
+    python pull_weights.py --towers reasoner,world \
+        --runtime-module reference_generator_runtime.py \
+        --gcs-dest gs://my-bucket/cosmos3
+
+Use --dry-run to print the file plan without touching disk or GCS.
+"""
+
+import argparse
+import fnmatch
+import os
+import shutil
+import subprocess
+import sys
+
+HF_REPO = "nvidia/Cosmos3-Edge"
+RUNTIME_MODULE_NAME = "cosmos3_generator_runtime.py"
+
+SNAPSHOT_IGNORE = ["assets/*", "images/*"]
+
+# Files each tower's package needs, as fnmatch patterns over repo-relative paths.
+TOWER_FILES = {
+    "reasoner": [
+        "config.json",
+        "generation_config.json",
+        "model.safetensors.index.json",
+        "chat_template.jinja",
+        "tokenizer.json",
+        "tokenizer_config.json",
+        "special_tokens_map.json",
+        "preprocessor_config.json",
+        "processor_config.json",
+        "video_preprocessor_config.json",
+        "vision_encoder/*",
+        # shared MoT weights - the reasoner index maps into these shards
+        "transformer/diffusion_pytorch_model*.safetensors",
+    ],
+    "world": [
+        "model_index.json",
+        "modular_model_index.json",
+        "negative_prompt.json",
+        "scheduler/*",
+        "transformer/*",
+        "vae/*",
+        "text_tokenizer/*",
+    ],
+}
+
+PACKAGE_NAMES = {
+    "reasoner": "cosmos-3-edge",
+    "world": "cosmos-3-edge-world",
+}
+
+
+def main() -> int:
+    args = _parse_args()
+    towers = [t.strip() for t in args.towers.split(",") if t.strip()]
+    unknown = set(towers).difference(TOWER_FILES)
+    if unknown:
+        raise SystemExit(f"Unknown tower(s): {sorted(unknown)}; pick from {sorted(TOWER_FILES)}")
+
+    snapshot_dir = args.snapshot_dir or _download_snapshot(args)
+    available = _list_snapshot_files(snapshot_dir)
+
+    for tower in towers:
+        package_dir = os.path.join(args.output_dir, PACKAGE_NAMES[tower])
+        selected = _select_files(available, TOWER_FILES[tower])
+        if not selected:
+            raise SystemExit(f"No files matched for tower '{tower}' in {snapshot_dir}")
+        print(f"[{tower}] {len(selected)} files -> {package_dir}")
+        if args.dry_run:
+            for rel in selected:
+                print(f"  {rel}")
+        else:
+            _materialize(snapshot_dir, package_dir, selected)
+        if tower == "world":
+            _inject_runtime_module(args, package_dir)
+        if args.gcs_dest:
+            _upload_to_gcs(
+                package_dir=package_dir,
+                gcs_dest=f"{args.gcs_dest.rstrip('/')}/{PACKAGE_NAMES[tower]}",
+                dry_run=args.dry_run,
+            )
+    return 0
+
+
+def _parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description=__doc__.split("\n")[0])
+    parser.add_argument("--hf-repo", default=HF_REPO)
+    parser.add_argument("--revision", default=None, help="HF revision/tag to pin.")
+    parser.add_argument(
+        "--snapshot-dir",
+        default=None,
+        help="Reuse an existing snapshot directory instead of downloading.",
+    )
+    parser.add_argument(
+        "--output-dir",
+        default="checkpoints/packages",
+        help="Directory receiving one package subdirectory per tower.",
+    )
+    parser.add_argument(
+        "--towers",
+        default="reasoner,world",
+        help="Comma-separated subset of: reasoner, world.",
+    )
+    parser.add_argument(
+        "--runtime-module",
+        default=None,
+        help=f"Python file copied into the world package as {RUNTIME_MODULE_NAME}.",
+    )
+    parser.add_argument(
+        "--gcs-dest",
+        default=None,
+        help="gs://bucket/prefix - each package is mirrored to <dest>/<package-name>/.",
+    )
+    parser.add_argument("--dry-run", action="store_true")
+    return parser.parse_args()
+
+
+def _download_snapshot(args: argparse.Namespace) -> str:
+    from huggingface_hub import snapshot_download
+
+    print(f"Downloading snapshot of {args.hf_repo} (weights only)...")
+    return snapshot_download(
+        args.hf_repo,
+        revision=args.revision,
+        ignore_patterns=SNAPSHOT_IGNORE,
+    )
+
+
+def _list_snapshot_files(snapshot_dir: str) -> list:
+    collected = []
+    for root, _, files in os.walk(snapshot_dir):
+        for name in files:
+            path = os.path.join(root, name)
+            rel = os.path.relpath(path, snapshot_dir)
+            if rel.startswith(".cache"):
+                continue
+            collected.append(rel)
+    return sorted(collected)
+
+
+def _select_files(available: list, patterns: list) -> list:
+    return [
+        rel
+        for rel in available
+        if any(fnmatch.fnmatch(rel, pattern) for pattern in patterns)
+    ]
+
+
+def _materialize(snapshot_dir: str, package_dir: str, selected: list) -> None:
+    for rel in selected:
+        src = os.path.realpath(os.path.join(snapshot_dir, rel))
+        dst = os.path.join(package_dir, rel)
+        os.makedirs(os.path.dirname(dst), exist_ok=True)
+        if os.path.exists(dst):
+            continue
+        try:
+            os.link(src, dst)
+        except OSError:
+            shutil.copy2(src, dst)
+
+
+def _inject_runtime_module(args: argparse.Namespace, package_dir: str) -> None:
+    if not args.runtime_module:
+        print(
+            f"  note: world package has no {RUNTIME_MODULE_NAME}; "
+            "Cosmos3EdgeWorldModel.from_pretrained will refuse to load it "
+            "(pass --runtime-module)."
+        )
+        return
+    dst = os.path.join(package_dir, RUNTIME_MODULE_NAME)
+    print(f"  runtime module: {args.runtime_module} -> {dst}")
+    if not args.dry_run:
+        os.makedirs(package_dir, exist_ok=True)
+        shutil.copy2(args.runtime_module, dst)
+
+
+def _upload_to_gcs(package_dir: str, gcs_dest: str, dry_run: bool) -> None:
+    tool = shutil.which("gcloud")
+    if tool:
+        cmd = ["gcloud", "storage", "cp", "-r", f"{package_dir.rstrip('/')}/*", gcs_dest + "/"]
+    elif shutil.which("gsutil"):
+        cmd = ["gsutil", "-m", "cp", "-r", f"{package_dir.rstrip('/')}/*", gcs_dest + "/"]
+    else:
+        raise SystemExit("Neither gcloud nor gsutil found on PATH for --gcs-dest.")
+    print(f"  upload: {' '.join(cmd)}")
+    if not dry_run:
+        subprocess.run(" ".join(cmd), shell=True, check=True)
+
+
+if __name__ == "__main__":
+    sys.exit(main())
diff --git a/inference_models/development/cosmos3/reference_generator_runtime.py b/inference_models/development/cosmos3/reference_generator_runtime.py
new file mode 100644
index 0000000000..7491faee87
--- /dev/null
+++ b/inference_models/development/cosmos3/reference_generator_runtime.py
@@ -0,0 +1,188 @@
+"""Reference `cosmos3_generator_runtime.py` for the Cosmos 3 Edge world package.
+
+This is the self-contained runtime module that ships INSIDE the
+`cosmos-3-edge-world` model package (injected by pull_weights.py
+--runtime-module). `Cosmos3EdgeWorldModel.from_pretrained` imports the
+`Cosmos3GeneratorRuntime` class from it via `import_class_from_file`.
+
+Backed by diffusers' `Cosmos3OmniPipeline` (diffusers>=0.39), so the only
+runtime dependencies are torch + diffusers + transformers - NVIDIA's
+cosmos-framework is not required for image-to-video.
+
+Contract expected by Cosmos3EdgeWorldModel (all images RGB numpy arrays):
+- load(package_dir, device) -> runtime
+- generate_video(images, prompt, num_frames, fps, resolution, seed) -> [frames]
+- encode_context(frames) -> session
+- rollout(session, actions, num_frames) -> ([frames], session)
+- infer_actions(frames) -> {actions, action_space, fps, metadata}
+
+Forward/inverse dynamics ride the pipeline's action-conditioning path
+(`CosmosActionCondition`): forward dynamics conditions each chunk on the last
+generated frame of the previous chunk (carried in the session dict), inverse
+dynamics infers the action sequence connecting the frames of an observation
+video. `action_space` maps to the pipeline's embodiment `domain_name`
+(e.g. "umi", "av", "droid_lerobot").
+"""
+
+from typing import List, Optional
+
+import numpy as np
+import torch
+
+
+class Cosmos3GeneratorRuntime:
+
+    @classmethod
+    def load(cls, package_dir: str, device: torch.device) -> "Cosmos3GeneratorRuntime":
+        from diffusers import Cosmos3OmniPipeline
+
+        pipeline = Cosmos3OmniPipeline.from_pretrained(
+            package_dir,
+            torch_dtype=torch.bfloat16,
+        )
+        pipeline.to(device)
+        return cls(pipeline=pipeline, device=device)
+
+    def __init__(self, pipeline, device: torch.device):
+        self._pipeline = pipeline
+        self._device = device
+
+    def generate_video(
+        self,
+        images: List[np.ndarray],
+        prompt: Optional[str],
+        num_frames: int,
+        fps: int,
+        resolution: int,
+        seed: Optional[int],
+        num_inference_steps: int = 35,
+    ) -> List[np.ndarray]:
+        from PIL import Image
+
+        conditioning = Image.fromarray(images[0])
+        generator = None
+        if seed is not None:
+            generator = torch.Generator(device=self._device).manual_seed(seed)
+        height, width = _resolve_resolution(resolution)
+        result = self._pipeline(
+            prompt=prompt or "",
+            image=conditioning,
+            num_frames=num_frames,
+            height=height,
+            width=width,
+            fps=float(fps),
+            num_inference_steps=num_inference_steps,
+            generator=generator,
+            output_type="np",
+            enable_safety_check=False,
+        )
+        return [frame for frame in _to_uint8(result.video)]
+
+    def encode_context(
+        self,
+        frames: List[np.ndarray],
+        prompt: Optional[str] = None,
+        view_point: str = "ego_view",
+        resolution_tier: int = 256,
+        **kwargs,
+    ) -> dict:
+        # The session carries the frame each subsequent chunk conditions on,
+        # plus the generation context that stays fixed across the rollout.
+        return {
+            "conditioning_frame": np.asarray(frames[-1]),
+            "prompt": prompt,
+            "view_point": view_point,
+            "resolution_tier": resolution_tier,
+        }
+
+    def rollout(
+        self,
+        session: dict,
+        actions: np.ndarray,
+        num_frames: int,
+        action_space: str = "umi",
+        seed: Optional[int] = None,
+        num_inference_steps: int = 35,
+        **kwargs,
+    ):
+        from diffusers.pipelines.cosmos.pipeline_cosmos3_omni import (
+            CosmosActionCondition,
+        )
+        from PIL import Image
+
+        condition = CosmosActionCondition(
+            mode="forward_dynamics",
+            chunk_size=len(actions),
+            domain_name=action_space,
+            resolution_tier=session["resolution_tier"],
+            raw_actions=torch.as_tensor(np.asarray(actions), dtype=torch.float32),
+            image=Image.fromarray(session["conditioning_frame"]),
+            view_point=session["view_point"],
+        )
+        generator = None
+        if seed is not None:
+            generator = torch.Generator(device=self._device).manual_seed(seed)
+        result = self._pipeline(
+            prompt=session.get("prompt") or "",
+            action=condition,
+            num_inference_steps=num_inference_steps,
+            generator=generator,
+            output_type="np",
+            enable_safety_check=False,
+        )
+        video = _to_uint8(result.video)
+        updated_session = dict(session)
+        updated_session["conditioning_frame"] = video[-1]
+        return [frame for frame in video], updated_session
+
+    def infer_actions(
+        self,
+        frames: List[np.ndarray],
+        action_space: str = "umi",
+        view_point: str = "ego_view",
+        resolution_tier: int = 256,
+        fps: float = 20.0,
+        num_inference_steps: int = 35,
+        **kwargs,
+    ) -> dict:
+        from diffusers.pipelines.cosmos.pipeline_cosmos3_omni import (
+            CosmosActionCondition,
+        )
+
+        condition = CosmosActionCondition(
+            mode="inverse_dynamics",
+            chunk_size=max(1, len(frames) - 1),
+            domain_name=action_space,
+            resolution_tier=resolution_tier,
+            video=[np.asarray(frame) for frame in frames],
+            view_point=view_point,
+        )
+        result = self._pipeline(
+            prompt="",
+            action=condition,
+            num_inference_steps=num_inference_steps,
+            output_type="np",
+            enable_safety_check=False,
+        )
+        chunks = [chunk.float().cpu().numpy() for chunk in (result.action or [])]
+        actions = np.concatenate(chunks, axis=0) if chunks else np.zeros((0, 0))
+        return {
+            "actions": actions,
+            "action_space": action_space,
+            "fps": fps,
+            "metadata": {"view_point": view_point, "num_chunks": len(chunks)},
+        }
+
+
+def _to_uint8(video) -> np.ndarray:
+    video = np.asarray(video)
+    if video.dtype != np.uint8:
+        video = (np.clip(video, 0.0, 1.0) * 255).astype(np.uint8)
+    return video
+
+
+def _resolve_resolution(resolution: int):
+    # 16:9 at the requested height, aligned to the VAE's spatial stride.
+    height = max(16, (resolution // 16) * 16)
+    width = max(16, ((resolution * 16 // 9) // 16) * 16)
+    return height, width
diff --git a/inference_models/docs/changelog.md b/inference_models/docs/changelog.md
index 325905365a..e7a96d1b6f 100644
--- a/inference_models/docs/changelog.md
+++ b/inference_models/docs/changelog.md
@@ -39,6 +39,15 @@ Add user-facing changes below using `### Added`, `### Changed`, `### Fixed`, or
 - Composable RF-DETR TensorRT execution plans, implementation contracts and registries,
   compatibility-aware implementation selection, and runtime metadata reporting the
   requested and effective preprocessing and postprocessing implementations.
+- NVIDIA Cosmos 3 Edge reasoner (`cosmos-3-edge`, task `vlm`, backend `hugging-face`):
+  image/video + text prompting via `prompt(...)` / `prompt_video(...)`, following the
+  standard VLM contract. The generative world-model tower ships separately.
+- NVIDIA Cosmos 3 Edge generator (`cosmos-3-edge-world`, task `world-model`, backend
+  `custom`): image-to-video (`generate_video`), forward dynamics (`start_rollout` +
+  `forward_dynamics` with explicit session-state threading), and inverse dynamics
+  (`inverse_dynamics`). The step-wise robot policy mode is deferred. The denoising
+  runtime ships inside the model package (loaded via `import_class_from_file`), keeping
+  NVIDIA's cosmos stack out of `inference_models` dependencies.
 - `segment_with_text_prompts` accepts `max_detections` (top-k by score, applied before mask
   interpolation; default `-1` = uncapped) and `mask_format` (`"dense"` default, or `"rle"`
   for COCO RLE at original resolution).
diff --git a/inference_models/inference_models/configuration.py b/inference_models/inference_models/configuration.py
index 3adbe6ccb7..9bbde56c47 100644
--- a/inference_models/inference_models/configuration.py
+++ b/inference_models/inference_models/configuration.py
@@ -237,6 +237,14 @@
     variable_name="INFERENCE_MODELS_QWEN3_VL_DEFAULT_DO_SAMPLE",
     default=INFERENCE_MODELS_DEFAULT_DO_SAMPLE,
 )
+INFERENCE_MODELS_COSMOS3_DEFAULT_MAX_NEW_TOKENS = get_integer_from_env(
+    variable_name="INFERENCE_MODELS_COSMOS3_DEFAULT_MAX_NEW_TOKENS",
+    default=512,
+)
+INFERENCE_MODELS_COSMOS3_DEFAULT_DO_SAMPLE = get_boolean_from_env(
+    variable_name="INFERENCE_MODELS_COSMOS3_DEFAULT_DO_SAMPLE",
+    default=INFERENCE_MODELS_DEFAULT_DO_SAMPLE,
+)
 INFERENCE_MODELS_GLM_OCR_DEFAULT_MAX_NEW_TOKENS = get_integer_from_env(
     variable_name="INFERENCE_MODELS_GLM_OCR_DEFAULT_MAX_NEW_TOKENS",
     default=8192,
diff --git a/inference_models/inference_models/models/auto_loaders/models_registry.py b/inference_models/inference_models/models/auto_loaders/models_registry.py
index 48345e3288..ccce3ae0df 100644
--- a/inference_models/inference_models/models/auto_loaders/models_registry.py
+++ b/inference_models/inference_models/models/auto_loaders/models_registry.py
@@ -23,6 +23,7 @@
 GAZE_DETECTION_TASK = "gaze-detection"
 OPEN_VOCABULARY_OBJECT_DETECTION_TASK = "open-vocabulary-object-detection"
 INTERACTIVE_INSTANCE_SEGMENTATION_TASK = "interactive-instance-segmentation"
+WORLD_MODEL_TASK = "world-model"
 
 
 @dataclass(frozen=True)
@@ -294,6 +295,14 @@ class RegistryEntry:
         module_name="inference_models.models.qwen3vl.qwen3vl_hf",
         class_name="Qwen3VLHF",
     ),
+    ("cosmos-3-edge", VLM_TASK, BackendType.HF): LazyClass(
+        module_name="inference_models.models.cosmos3.cosmos3_reasoner_hf",
+        class_name="Cosmos3EdgeReasoner",
+    ),
+    ("cosmos-3-edge-world", WORLD_MODEL_TASK, BackendType.CUSTOM): LazyClass(
+        module_name="inference_models.models.cosmos3.cosmos3_world",
+        class_name="Cosmos3EdgeWorldModel",
+    ),
     ("qwen3_5", VLM_TASK, BackendType.HF): LazyClass(
         module_name="inference_models.models.qwen3_5.qwen3_5_hf",
         class_name="Qwen35HF",
diff --git a/inference_models/inference_models/models/cosmos3/LICENSE.txt b/inference_models/inference_models/models/cosmos3/LICENSE.txt
new file mode 100644
index 0000000000..52f32cc9c1
--- /dev/null
+++ b/inference_models/inference_models/models/cosmos3/LICENSE.txt
@@ -0,0 +1,55 @@
+NVIDIA Cosmos 3 Edge is released under the OpenMDW License Agreement,
+version 1.1 (OpenMDW-1.1): https://openmdw.ai/license/1-1/
+
+----------------------------------------------------------------------
+
+OpenMDW License Agreement, version 1.1 (OpenMDW-1.1)
+
+By exercising rights granted to you under this agreement, you accept and
+agree to its terms.
+
+As used in this agreement, "Model Materials" means the materials provided
+to you under this agreement, consisting of: (1) one or more machine
+learning models (including architecture and parameters); and (2) all
+related artifacts (including associated data, documentation and software)
+that are provided to you hereunder.
+
+Subject to your compliance with this agreement, permission is hereby
+granted, free of charge, to deal in the Model Materials without
+restriction, including under all copyright, patent, database, and trade
+secret rights included or embodied therein.
+
+If you distribute any portion of the Model Materials, you shall retain in
+your distribution (1) a copy of this agreement, and (2) all copyright
+notices and other notices of origin included in the Model Materials that
+are applicable to your distribution.
+
+If you file, maintain, or voluntarily participate in a lawsuit against
+any person or entity asserting that the Model Materials directly or
+indirectly infringe any patent or copyright, then all rights and grants
+made to you hereunder are terminated, unless that lawsuit was in response
+to a corresponding lawsuit first brought against you.
+
+This agreement does not impose any restrictions or obligations with
+respect to any use, modification, or sharing of any outputs generated by
+using the Model Materials.
+
+THE MODEL MATERIALS ARE PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, TITLE, NONINFRINGEMENT,
+ACCURACY, OR THE ABSENCE OF LATENT OR OTHER DEFECTS OR ERRORS, WHETHER OR
+NOT DISCOVERABLE, ALL TO THE GREATEST EXTENT PERMISSIBLE UNDER APPLICABLE
+LAW.
+
+YOU ARE SOLELY RESPONSIBLE FOR (1) CLEARING RIGHTS OF OTHER PERSONS THAT
+MAY APPLY TO THE MODEL MATERIALS OR ANY USE THEREOF, INCLUDING WITHOUT
+LIMITATION ANY PERSON'S COPYRIGHTS OR OTHER RIGHTS INCLUDED OR EMBODIED IN
+THE MODEL MATERIALS; (2) OBTAINING ANY NECESSARY CONSENTS, PERMISSIONS OR
+OTHER RIGHTS REQUIRED FOR ANY USE OF THE MODEL MATERIALS; OR (3)
+PERFORMING ANY DUE DILIGENCE OR UNDERTAKING ANY OTHER INVESTIGATIONS INTO
+THE MODEL MATERIALS OR ANYTHING INCORPORATED OR EMBODIED THEREIN.
+
+IN NO EVENT SHALL THE PROVIDERS OF THE MODEL MATERIALS BE LIABLE FOR ANY
+CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT
+OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE MODEL
+MATERIALS, THE USE THEREOF OR OTHER DEALINGS THEREIN.
diff --git a/inference_models/inference_models/models/cosmos3/__init__.py b/inference_models/inference_models/models/cosmos3/__init__.py
new file mode 100644
index 0000000000..e69de29bb2
diff --git a/inference_models/inference_models/models/cosmos3/cosmos3_reasoner_hf.py b/inference_models/inference_models/models/cosmos3/cosmos3_reasoner_hf.py
new file mode 100644
index 0000000000..3e3ef18ecc
--- /dev/null
+++ b/inference_models/inference_models/models/cosmos3/cosmos3_reasoner_hf.py
@@ -0,0 +1,295 @@
+"""
+This is inference-models wrapper for the reasoner tower of NVIDIA Cosmos 3 Edge,
+originally published in https://huggingface.co/nvidia/Cosmos3-Edge
+"""
+
+import re
+from threading import Lock
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+from transformers import AutoModelForImageTextToText, AutoProcessor
+from transformers.utils import is_flash_attn_2_available
+
+from inference_models.configuration import (
+    DEFAULT_DEVICE,
+    INFERENCE_MODELS_COSMOS3_DEFAULT_DO_SAMPLE,
+    INFERENCE_MODELS_COSMOS3_DEFAULT_MAX_NEW_TOKENS,
+    RUNNING_ON_JETSON,
+)
+from inference_models.entities import ColorFormat
+
+DEFAULT_PROMPT = "Describe what's in this image."
+SYSTEM_PROMPT_SENTINEL = "<system_prompt>"
+THINK_BLOCK_PATTERN = re.compile(r"<think>.*?</think>\s*", flags=re.DOTALL)
+THINK_EXTRACT_PATTERN = re.compile(r"<think>(.*?)</think>", flags=re.DOTALL)
+
+
+def _get_cosmos3_attn_implementation(device: torch.device) -> str:
+    if (
+        is_flash_attn_2_available()
+        and device
+        and "cuda" in str(device)
+        and not RUNNING_ON_JETSON
+    ):
+        try:
+            import flash_attn  # noqa: F401
+
+            if _is_ampere_plus(device=device):
+                return "flash_attention_2"
+            return "eager"
+        except ImportError:
+            pass
+    return "eager"
+
+
+def _is_ampere_plus(device: torch.device) -> bool:
+    if device.type != "cuda":
+        return False
+    major, _ = torch.cuda.get_device_capability(device=device)
+    return major >= 8
+
+
+def _resolve_default_dtype(device: torch.device) -> torch.dtype:
+    if device.type == "cuda":
+        if torch.cuda.is_bf16_supported():
+            return torch.bfloat16
+        return torch.float16
+    return torch.float32
+
+
+class Cosmos3EdgeReasoner:
+    """NVIDIA Cosmos 3 Edge reasoner tower (image/video + text -> text).
+
+    Only the autoregressive reasoner is exposed here; the diffusion generator
+    (image-to-video, dynamics, policy) has a separate implementation and
+    registry entry.
+    """
+
+    default_dtype = torch.bfloat16
+
+    @classmethod
+    def from_pretrained(
+        cls,
+        model_name_or_path: str,
+        device: torch.device = DEFAULT_DEVICE,
+        trust_remote_code: bool = False,
+        local_files_only: bool = True,
+        quantization_config: Any = None,
+        **kwargs,
+    ) -> "Cosmos3EdgeReasoner":
+        dtype = _resolve_default_dtype(device)
+        attn_implementation = _get_cosmos3_attn_implementation(device)
+        model = AutoModelForImageTextToText.from_pretrained(
+            model_name_or_path,
+            device_map=device,
+            dtype=dtype,
+            trust_remote_code=trust_remote_code,
+            local_files_only=local_files_only,
+            quantization_config=quantization_config,
+            attn_implementation=attn_implementation,
+        ).eval()
+        processor = AutoProcessor.from_pretrained(
+            model_name_or_path,
+            trust_remote_code=trust_remote_code,
+            local_files_only=local_files_only,
+        )
+        return cls(model=model, processor=processor, device=device)
+
+    def __init__(
+        self,
+        model,
+        processor,
+        device: torch.device,
+    ):
+        self._model = model
+        self._processor = processor
+        self._device = device
+        self._torch_dtype = next(model.parameters()).dtype
+        self.default_system_prompt = (
+            "You are Cosmos, a helpful assistant that understands physical scenes "
+            "and answers questions about images and videos."
+        )
+        self._lock = Lock()
+
+    def prompt(
+        self,
+        images: Union[torch.Tensor, List[torch.Tensor], np.ndarray, List[np.ndarray]],
+        prompt: str = None,
+        input_color_format: ColorFormat = None,
+        max_new_tokens: Optional[int] = INFERENCE_MODELS_COSMOS3_DEFAULT_MAX_NEW_TOKENS,
+        do_sample: bool = INFERENCE_MODELS_COSMOS3_DEFAULT_DO_SAMPLE,
+        skip_special_tokens: bool = True,
+        return_thinking: bool = False,
+        **kwargs,
+    ) -> Union[List[str], List[Dict[str, str]]]:
+        inputs = self.pre_process_generation(
+            images=images, prompt=prompt, input_color_format=input_color_format
+        )
+        generated_ids = self.generate(
+            inputs=inputs,
+            max_new_tokens=max_new_tokens,
+            do_sample=do_sample,
+        )
+        return self.post_process_generation(
+            generated_ids=generated_ids,
+            skip_special_tokens=skip_special_tokens,
+            return_thinking=return_thinking,
+        )
+
+    def prompt_video(
+        self,
+        frames: List[np.ndarray],
+        prompt: str = None,
+        input_color_format: ColorFormat = None,
+        max_new_tokens: Optional[int] = INFERENCE_MODELS_COSMOS3_DEFAULT_MAX_NEW_TOKENS,
+        do_sample: bool = INFERENCE_MODELS_COSMOS3_DEFAULT_DO_SAMPLE,
+        skip_special_tokens: bool = True,
+        return_thinking: bool = False,
+        **kwargs,
+    ) -> Union[str, Dict[str, str]]:
+        inputs = self.pre_process_generation(
+            images=frames,
+            prompt=prompt,
+            input_color_format=input_color_format,
+            as_video=True,
+        )
+        generated_ids = self.generate(
+            inputs=inputs,
+            max_new_tokens=max_new_tokens,
+            do_sample=do_sample,
+        )
+        return self.post_process_generation(
+            generated_ids=generated_ids,
+            skip_special_tokens=skip_special_tokens,
+            return_thinking=return_thinking,
+        )[0]
+
+    def pre_process_generation(
+        self,
+        images: Union[torch.Tensor, List[torch.Tensor], np.ndarray, List[np.ndarray]],
+        prompt: str = None,
+        input_color_format: ColorFormat = None,
+        as_video: bool = False,
+        **kwargs,
+    ) -> dict:
+        if isinstance(images, np.ndarray):
+            if input_color_format != "rgb":
+                images = images[:, :, ::-1]
+            images = images.copy()
+        elif as_video and input_color_format != "rgb":
+            images = [
+                frame[:, :, ::-1].copy() if isinstance(frame, np.ndarray) else frame
+                for frame in images
+            ]
+        prompt, system_prompt = self._parse_prompt(prompt=prompt)
+        visual_content = (
+            {"type": "video", "video": images}
+            if as_video
+            else {"type": "image", "image": images}
+        )
+        conversation = [
+            {
+                "role": "system",
+                "content": [{"type": "text", "text": system_prompt}],
+            },
+            {
+                "role": "user",
+                "content": [
+                    visual_content,
+                    {"type": "text", "text": prompt},
+                ],
+            },
+        ]
+        text_input = self._processor.apply_chat_template(
+            conversation, tokenize=False, add_generation_prompt=True
+        )
+        processor_kwargs = {"videos": [images]} if as_video else {"images": images}
+        model_inputs = self._processor(
+            text=text_input,
+            return_tensors="pt",
+            padding=True,
+            **processor_kwargs,
+        )
+        return {
+            k: (
+                v.to(self._device, dtype=self._torch_dtype)
+                if v.is_floating_point()
+                else v.to(self._device)
+            )
+            for k, v in model_inputs.items()
+            if isinstance(v, torch.Tensor)
+        }
+
+    def generate(
+        self,
+        inputs: dict,
+        max_new_tokens: Optional[int] = INFERENCE_MODELS_COSMOS3_DEFAULT_MAX_NEW_TOKENS,
+        do_sample: bool = INFERENCE_MODELS_COSMOS3_DEFAULT_DO_SAMPLE,
+        **kwargs,
+    ) -> torch.Tensor:
+        if max_new_tokens is None:
+            max_new_tokens = INFERENCE_MODELS_COSMOS3_DEFAULT_MAX_NEW_TOKENS
+        input_len = inputs["input_ids"].shape[-1]
+        tokenizer = self._processor.tokenizer
+        pad_token_id = (
+            getattr(tokenizer, "pad_token_id", None) or tokenizer.eos_token_id
+        )
+        with self._lock, torch.inference_mode():
+            generation = self._model.generate(
+                **inputs,
+                max_new_tokens=max_new_tokens,
+                do_sample=do_sample,
+                pad_token_id=pad_token_id,
+                eos_token_id=tokenizer.eos_token_id,
+            )
+        return generation[:, input_len:]
+
+    def post_process_generation(
+        self,
+        generated_ids: torch.Tensor,
+        skip_special_tokens: bool = True,
+        return_thinking: bool = False,
+        **kwargs,
+    ) -> Union[List[str], List[Dict[str, str]]]:
+        decoded = self._processor.batch_decode(
+            generated_ids,
+            skip_special_tokens=skip_special_tokens,
+        )
+        result = []
+        for text in decoded:
+            text = text.replace("assistant\n", "")
+            # The chat template opens the reasoning block inside the prompt, so
+            # decoded output carries a bare closing </think>. Restore the
+            # opening tag so thinking and answer parse apart (qwen3_5 pattern).
+            if "</think>" in text and "<think>" not in text:
+                text = "<think>" + text
+            if return_thinking:
+                think_match = THINK_EXTRACT_PATTERN.search(text)
+                if think_match:
+                    thinking = think_match.group(1).strip()
+                    answer = THINK_BLOCK_PATTERN.sub("", text).strip()
+                else:
+                    thinking = text.replace("<think>", "").strip()
+                    answer = ""
+                result.append({"thinking": thinking, "answer": answer})
+            else:
+                result.append(THINK_BLOCK_PATTERN.sub("", text).strip())
+        return result
+
+    def __call__(
+        self,
+        images: Union[torch.Tensor, List[torch.Tensor], np.ndarray, List[np.ndarray]],
+        **kwargs,
+    ) -> Union[List[str], List[Dict[str, str]]]:
+        return self.prompt(images, **kwargs)
+
+    def _parse_prompt(self, prompt: Optional[str]) -> Tuple[str, str]:
+        if prompt is None:
+            return DEFAULT_PROMPT, self.default_system_prompt
+        split_prompt = prompt.split(SYSTEM_PROMPT_SENTINEL)
+        parsed_prompt = split_prompt[0] or DEFAULT_PROMPT
+        if len(split_prompt) == 1:
+            return parsed_prompt, self.default_system_prompt
+        return parsed_prompt, split_prompt[1] or self.default_system_prompt
diff --git a/inference_models/inference_models/models/cosmos3/cosmos3_world.py b/inference_models/inference_models/models/cosmos3/cosmos3_world.py
new file mode 100644
index 0000000000..ed366b2e3a
--- /dev/null
+++ b/inference_models/inference_models/models/cosmos3/cosmos3_world.py
@@ -0,0 +1,213 @@
+from dataclasses import dataclass, field
+from threading import Lock
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from inference_models.configuration import DEFAULT_DEVICE
+from inference_models.entities import ColorFormat
+from inference_models.errors import ModelInputError
+from inference_models.models.cosmos3.runtime_loading import load_runtime_from_package
+
+RUNTIME_MODULE_FILE = "cosmos3_generator_runtime.py"
+RUNTIME_CLASS_NAME = "Cosmos3GeneratorRuntime"
+SESSION_KEY = "cosmos3_rollout_session"
+
+DEFAULT_NUM_FRAMES = 121
+DEFAULT_FPS = 24
+DEFAULT_RESOLUTION = 480
+
+
+@dataclass
+class Cosmos3ActionTrajectory:
+    """A robot action trajectory - consumed by forward dynamics, produced by
+    inverse dynamics."""
+
+    actions: np.ndarray  # (T, action_dim)
+    action_space: str
+    fps: float
+    metadata: dict = field(default_factory=dict)
+
+
+@dataclass
+class Cosmos3Rollout:
+    """Result of a generation / dynamics call.
+
+    `frames` are BGR (repo-wide convention). `state_dict` is opaque live
+    session state - thread it back into the next `forward_dynamics()` call,
+    never introspect or serialize it.
+    """
+
+    frames: List[np.ndarray]
+    fps: float
+    resolution: Tuple[int, int]
+    actions: Optional[Cosmos3ActionTrajectory] = None
+    state_dict: Optional[dict] = None
+
+
+class Cosmos3EdgeWorldModel:
+    """NVIDIA Cosmos 3 Edge generator tower.
+
+    Modes exposed: image-to-video (`generate_video`), forward dynamics
+    (`start_rollout` + `forward_dynamics`), inverse dynamics
+    (`inverse_dynamics`). The policy mode (step-wise robot control) is
+    deliberately not exposed yet.
+    """
+
+    @classmethod
+    def from_pretrained(
+        cls,
+        model_name_or_path: str,
+        device: torch.device = DEFAULT_DEVICE,
+        **kwargs,
+    ) -> "Cosmos3EdgeWorldModel":
+        runtime_class = load_runtime_from_package(
+            model_name_or_path=model_name_or_path,
+            runtime_module_file=RUNTIME_MODULE_FILE,
+            runtime_class_name=RUNTIME_CLASS_NAME,
+        )
+        runtime = runtime_class.load(model_name_or_path, device=device)
+        return cls(runtime=runtime, device=device)
+
+    def __init__(self, runtime, device: torch.device):
+        self._runtime = runtime
+        self._device = device
+        self._lock = Lock()
+
+    def generate_video(
+        self,
+        images: Union[np.ndarray, List[np.ndarray]],
+        prompt: Optional[str] = None,
+        num_frames: int = DEFAULT_NUM_FRAMES,
+        fps: int = DEFAULT_FPS,
+        resolution: int = DEFAULT_RESOLUTION,
+        seed: Optional[int] = None,
+        input_color_format: ColorFormat = None,
+        **kwargs,
+    ) -> Cosmos3Rollout:
+        conditioning = _as_rgb_frames(images, input_color_format=input_color_format)
+        with self._lock:
+            generated = self._runtime.generate_video(
+                images=conditioning,
+                prompt=prompt,
+                num_frames=num_frames,
+                fps=fps,
+                resolution=resolution,
+                seed=seed,
+            )
+        frames = [_rgb_to_bgr(frame) for frame in generated]
+        return Cosmos3Rollout(
+            frames=frames,
+            fps=float(fps),
+            resolution=_frames_resolution(frames),
+        )
+
+    def start_rollout(
+        self,
+        frames: List[np.ndarray],
+        input_color_format: ColorFormat = None,
+        **session_options,
+    ) -> dict:
+        context = _as_rgb_frames(frames, input_color_format=input_color_format)
+        if not context:
+            raise ModelInputError(
+                message="start_rollout() requires at least one context frame.",
+                help_url="https://inference-models.roboflow.com/errors/models-input/#modelinputerror",
+            )
+        with self._lock:
+            session = self._runtime.encode_context(frames=context, **session_options)
+        return {SESSION_KEY: session}
+
+    def forward_dynamics(
+        self,
+        actions: Cosmos3ActionTrajectory,
+        state_dict: dict,
+        num_frames: Optional[int] = None,
+        **kwargs,
+    ) -> Cosmos3Rollout:
+        session = self._resolve_session(state_dict=state_dict)
+        if num_frames is None:
+            num_frames = len(actions.actions)
+        with self._lock:
+            generated, updated_session = self._runtime.rollout(
+                session=session,
+                actions=actions.actions,
+                num_frames=num_frames,
+                action_space=actions.action_space,
+                **kwargs,
+            )
+        frames = [_rgb_to_bgr(frame) for frame in generated]
+        return Cosmos3Rollout(
+            frames=frames,
+            fps=actions.fps,
+            resolution=_frames_resolution(frames),
+            state_dict={SESSION_KEY: updated_session},
+        )
+
+    def inverse_dynamics(
+        self,
+        frames: List[np.ndarray],
+        action_space: str,
+        input_color_format: ColorFormat = None,
+        **kwargs,
+    ) -> Cosmos3ActionTrajectory:
+        observation = _as_rgb_frames(frames, input_color_format=input_color_format)
+        if not observation:
+            raise ModelInputError(
+                message="inverse_dynamics() requires at least one observation frame.",
+                help_url="https://inference-models.roboflow.com/errors/models-input/#modelinputerror",
+            )
+        with self._lock:
+            result = self._runtime.infer_actions(
+                frames=observation, action_space=action_space, **kwargs
+            )
+        return Cosmos3ActionTrajectory(
+            actions=np.asarray(result["actions"]),
+            action_space=result["action_space"],
+            fps=float(result["fps"]),
+            metadata=result.get("metadata", {}),
+        )
+
+    def _resolve_session(self, state_dict: dict):
+        # A missing / foreign state_dict must fail loudly - silently creating
+        # an empty session would produce plausible-looking garbage rollouts.
+        if not isinstance(state_dict, dict) or SESSION_KEY not in state_dict:
+            raise ModelInputError(
+                message="forward_dynamics() requires the state_dict returned by "
+                "start_rollout() (or by a previous forward_dynamics() call). "
+                "Call start_rollout() with context frames first.",
+                help_url="https://inference-models.roboflow.com/errors/models-input/#modelinputerror",
+            )
+        return state_dict[SESSION_KEY]
+
+
+def _as_rgb_frames(
+    images: Union[np.ndarray, List[np.ndarray]],
+    input_color_format: ColorFormat,
+) -> List[np.ndarray]:
+    if isinstance(images, np.ndarray):
+        images = [images]
+    frames = []
+    for frame in images:
+        if not isinstance(frame, np.ndarray):
+            raise ModelInputError(
+                message="Cosmos 3 generator inputs must be numpy arrays "
+                f"(got {type(frame).__name__}).",
+                help_url="https://inference-models.roboflow.com/errors/models-input/#modelinputerror",
+            )
+        if input_color_format != "rgb":
+            frame = frame[:, :, ::-1]
+        frames.append(np.ascontiguousarray(frame))
+    return frames
+
+
+def _rgb_to_bgr(frame: np.ndarray) -> np.ndarray:
+    return np.ascontiguousarray(np.asarray(frame)[:, :, ::-1])
+
+
+def _frames_resolution(frames: List[np.ndarray]) -> Tuple[int, int]:
+    if not frames:
+        return (0, 0)
+    height, width = frames[0].shape[:2]
+    return (width, height)
diff --git a/inference_models/inference_models/models/cosmos3/runtime_loading.py b/inference_models/inference_models/models/cosmos3/runtime_loading.py
new file mode 100644
index 0000000000..b6d54a1f03
--- /dev/null
+++ b/inference_models/inference_models/models/cosmos3/runtime_loading.py
@@ -0,0 +1,31 @@
+import os
+
+from inference_models.errors import ModelRuntimeError
+from inference_models.utils.imports import import_class_from_file
+
+
+def load_runtime_from_package(
+    model_name_or_path: str,
+    runtime_module_file: str,
+    runtime_class_name: str,
+) -> type:
+    """Load a generator runtime class shipped inside a model package.
+
+    The Cosmos 3 generator towers are not loadable through plain transformers:
+    their denoising loops come from NVIDIA's cosmos stack. Instead of making
+    that stack an `inference_models` dependency, the model package ships a
+    self-contained runtime module which this helper imports (same mechanism as
+    moondream2's package-local code).
+    """
+    runtime_path = os.path.join(model_name_or_path, runtime_module_file)
+    if not os.path.exists(runtime_path):
+        raise ModelRuntimeError(
+            message=f"Model package at {model_name_or_path} does not contain the "
+            f"expected runtime module `{runtime_module_file}`. The package may be "
+            "corrupted or predate generator support.",
+            help_url="https://inference-models.roboflow.com/errors/models-runtime/#modelruntimeerror",
+        )
+    return import_class_from_file(
+        file_path=runtime_path,
+        class_name=runtime_class_name,
+    )
diff --git a/inference_models/tests/integration_tests/models/test_cosmos3_predictions.py b/inference_models/tests/integration_tests/models/test_cosmos3_predictions.py
new file mode 100644
index 0000000000..20187910bc
--- /dev/null
+++ b/inference_models/tests/integration_tests/models/test_cosmos3_predictions.py
@@ -0,0 +1,114 @@
+"""Integration tests for Cosmos 3 Edge (reasoner + world model) on real weights.
+
+Both tests load from local package directories (the layouts produced by
+`development/cosmos3/pull_weights.py`) pointed to by env vars, so they can run
+before the packages are published to the weights provider:
+
+    COSMOS3_REASONER_PACKAGE_DIR=checkpoints/packages/cosmos-3-edge \
+    COSMOS3_WORLD_PACKAGE_DIR=checkpoints/packages/cosmos-3-edge-world \
+    python -m pytest tests/integration_tests/models/test_cosmos3_predictions.py -m slow
+
+Once the packages are registered, conftest fixtures downloading the published
+zips should replace the env-var indirection (matching the other model suites).
+"""
+
+import os
+
+import numpy as np
+import pytest
+import torch
+
+REASONER_PACKAGE_DIR = os.environ.get("COSMOS3_REASONER_PACKAGE_DIR")
+WORLD_PACKAGE_DIR = os.environ.get("COSMOS3_WORLD_PACKAGE_DIR")
+CUDA_AVAILABLE = torch.cuda.is_available()
+
+
+@pytest.mark.slow
+@pytest.mark.skipif(
+    not REASONER_PACKAGE_DIR or not CUDA_AVAILABLE,
+    reason="COSMOS3_REASONER_PACKAGE_DIR not set or CUDA unavailable",
+)
+def test_cosmos3_reasoner_answers_scene_question() -> None:
+    from inference_models.models.cosmos3.cosmos3_reasoner_hf import (
+        Cosmos3EdgeReasoner,
+    )
+
+    model = Cosmos3EdgeReasoner.from_pretrained(
+        REASONER_PACKAGE_DIR, device=torch.device("cuda")
+    )
+    image = np.zeros((240, 320, 3), dtype=np.uint8)
+    image[:, 160:, 2] = 255  # right half red (BGR)
+
+    answers = model.prompt(
+        images=image,
+        prompt="Which side of the image is red? Answer with 'left' or 'right'.",
+        max_new_tokens=64,
+    )
+
+    assert len(answers) == 1
+    assert isinstance(answers[0], str)
+    assert len(answers[0].strip()) > 0
+
+
+@pytest.mark.slow
+@pytest.mark.skipif(
+    not WORLD_PACKAGE_DIR or not CUDA_AVAILABLE,
+    reason="COSMOS3_WORLD_PACKAGE_DIR not set or CUDA unavailable",
+)
+def test_cosmos3_world_generates_video_from_image() -> None:
+    from inference_models.models.cosmos3.cosmos3_world import Cosmos3EdgeWorldModel
+
+    model = Cosmos3EdgeWorldModel.from_pretrained(
+        WORLD_PACKAGE_DIR, device=torch.device("cuda")
+    )
+    conditioning = np.full((240, 416, 3), 128, dtype=np.uint8)
+
+    rollout = model.generate_video(
+        images=conditioning,
+        prompt="static gray scene",
+        num_frames=5,
+        fps=24,
+        resolution=240,
+        seed=0,
+        num_inference_steps=10,
+    )
+
+    assert len(rollout.frames) >= 1
+    assert rollout.frames[0].ndim == 3
+    assert rollout.frames[0].shape[2] == 3
+    assert rollout.frames[0].dtype == np.uint8
+
+
+@pytest.mark.slow
+@pytest.mark.skipif(
+    not WORLD_PACKAGE_DIR or not CUDA_AVAILABLE,
+    reason="COSMOS3_WORLD_PACKAGE_DIR not set or CUDA unavailable",
+)
+def test_cosmos3_world_forward_dynamics_threads_session() -> None:
+    from inference_models.models.cosmos3.cosmos3_world import (
+        Cosmos3ActionTrajectory,
+        Cosmos3EdgeWorldModel,
+    )
+
+    model = Cosmos3EdgeWorldModel.from_pretrained(
+        WORLD_PACKAGE_DIR, device=torch.device("cuda")
+    )
+    first_frame = np.full((256, 256, 3), 128, dtype=np.uint8)
+    actions = Cosmos3ActionTrajectory(
+        actions=np.zeros((8, 10), dtype=np.float32),
+        action_space="umi",
+        fps=20.0,
+    )
+
+    state = model.start_rollout(frames=[first_frame], resolution_tier=256)
+    rollout = model.forward_dynamics(
+        actions=actions, state_dict=state, num_inference_steps=10
+    )
+
+    assert len(rollout.frames) >= 1
+    assert rollout.state_dict is not None
+    # the returned state must be usable for the next chunk
+    second = model.forward_dynamics(
+        actions=actions, state_dict=rollout.state_dict, num_inference_steps=10
+    )
+    assert len(second.frames) >= 1
diff --git a/inference_models/tests/unit_tests/models/test_cosmos3_reasoner_hf.py b/inference_models/tests/unit_tests/models/test_cosmos3_reasoner_hf.py
new file mode 100644
index 0000000000..bb99599e45
--- /dev/null
+++ b/inference_models/tests/unit_tests/models/test_cosmos3_reasoner_hf.py
@@ -0,0 +1,151 @@
+from unittest.mock import MagicMock
+
+import numpy as np
+import torch
+
+from inference_models.configuration import (
+    INFERENCE_MODELS_COSMOS3_DEFAULT_MAX_NEW_TOKENS,
+)
+from inference_models.models.cosmos3.cosmos3_reasoner_hf import Cosmos3EdgeReasoner
+
+
+def _model_with_processor() -> Cosmos3EdgeReasoner:
+    model = MagicMock()
+    model.parameters.return_value = iter([torch.tensor(0.0, dtype=torch.bfloat16)])
+    processor = MagicMock()
+    processor.apply_chat_template.return_value = "templated"
+    processor.return_value = {
+        "input_ids": torch.tensor([[1, 2, 3]], dtype=torch.int64),
+        "pixel_values": torch.zeros((1, 3, 8, 8), dtype=torch.float32),
+    }
+    return Cosmos3EdgeReasoner(
+        model=model, processor=processor, device=torch.device("cpu")
+    )
+
+
+def test_generate_returns_only_new_tokens() -> None:
+    reasoner = _model_with_processor()
+    reasoner._model.generate.return_value = torch.tensor([[1, 2, 21, 22]])
+
+    result = reasoner.generate(inputs={"input_ids": torch.tensor([[1, 2]])})
+
+    assert result.tolist() == [[21, 22]]
+
+
+def test_pre_process_generation_builds_system_and_user_turns() -> None:
+    reasoner = _model_with_processor()
+
+    inputs = reasoner.pre_process_generation(
+        images=np.zeros((8, 8, 3), dtype=np.uint8),
+        prompt="What is happening?<system_prompt>Be terse.",
+    )
+
+    conversation = reasoner._processor.apply_chat_template.call_args.args[0]
+    assert conversation[0]["role"] == "system"
+    assert conversation[0]["content"][0]["text"] == "Be terse."
+    assert conversation[1]["content"][1]["text"] == "What is happening?"
+    assert "input_ids" in inputs and "pixel_values" in inputs
+
+
+def test_pre_process_generation_uses_defaults_without_prompt() -> None:
+    reasoner = _model_with_processor()
+
+    reasoner.pre_process_generation(images=np.zeros((8, 8, 3), dtype=np.uint8))
+
+    conversation = reasoner._processor.apply_chat_template.call_args.args[0]
+    assert conversation[0]["content"][0]["text"] == reasoner.default_system_prompt
+    assert conversation[1]["content"][1]["text"] == "Describe what's in this image."
+
+
+def test_pre_process_generation_video_path_passes_frames_as_video() -> None:
+    reasoner = _model_with_processor()
+    frames = [np.zeros((8, 8, 3), dtype=np.uint8) for _ in range(4)]
+
+    reasoner.pre_process_generation(images=frames, as_video=True)
+
+    conversation = reasoner._processor.apply_chat_template.call_args.args[0]
+    assert conversation[1]["content"][0]["type"] == "video"
+    assert "videos" in reasoner._processor.call_args.kwargs
+    assert len(reasoner._processor.call_args.kwargs["videos"][0]) == 4
+
+
+def test_pre_process_generation_casts_floating_point_inputs_to_model_dtype() -> None:
+    reasoner = _model_with_processor()
+
+    inputs = reasoner.pre_process_generation(images=np.zeros((8, 8, 3), dtype=np.uint8))
+
+    assert inputs["input_ids"].dtype == torch.int64
+    assert inputs["pixel_values"].dtype == torch.bfloat16
+
+
+def test_generate_uses_default_max_new_tokens_when_none_is_given() -> None:
+    reasoner = _model_with_processor()
+    reasoner._model.generate.return_value = torch.tensor([[1, 2, 9]])
+
+    reasoner.generate(inputs={"input_ids": torch.tensor([[1, 2]])}, max_new_tokens=None)
+
+    assert reasoner._model.generate.call_args.kwargs["max_new_tokens"] == (
+        INFERENCE_MODELS_COSMOS3_DEFAULT_MAX_NEW_TOKENS
+    )
+
+
+def test_post_process_generation_strips_thinking_block_by_default() -> None:
+    reasoner = _model_with_processor()
+    reasoner._processor.batch_decode.return_value = [
+        "The image is blue. So the answer is blue.</think>  blue"
+    ]
+
+    result = reasoner.post_process_generation(generated_ids=torch.tensor([[1]]))
+
+    assert result == ["blue"]
+
+
+def test_post_process_generation_returns_thinking_when_requested() -> None:
+    reasoner = _model_with_processor()
+    reasoner._processor.batch_decode.return_value = [
+        "The image is blue. So the answer is blue.</think>  blue"
+    ]
+
+    result = reasoner.post_process_generation(
+        generated_ids=torch.tensor([[1]]), return_thinking=True
+    )
+
+    assert result == [
+        {
+            "thinking": "The image is blue. So the answer is blue.",
+            "answer": "blue",
+        }
+    ]
+
+
+def test_post_process_generation_handles_truncated_thinking() -> None:
+    reasoner = _model_with_processor()
+    reasoner._processor.batch_decode.return_value = ["Okay, the user wants"]
+
+    result = reasoner.post_process_generation(
+        generated_ids=torch.tensor([[1]]), return_thinking=True
+    )
+
+    assert result == [{"thinking": "Okay, the user wants", "answer": ""}]
+
+
+def test_post_process_generation_strips_assistant_prefix() -> None:
+    reasoner = _model_with_processor()
+    reasoner._processor.batch_decode.return_value = ["assistant\nThe box falls.  "]
+
+    result = reasoner.post_process_generation(generated_ids=torch.tensor([[1]]))
+
+    assert result == ["The box falls."]
+
+
+def test_prompt_video_returns_single_string() -> None:
+    reasoner = _model_with_processor()
+    reasoner._model.generate.return_value = torch.tensor([[1, 2, 3, 9]])
+    reasoner._processor.batch_decode.return_value = ["a robot arm"]
+
+    result = reasoner.prompt_video(
+        frames=[np.zeros((8, 8, 3), dtype=np.uint8)] * 2,
+        prompt="What will happen next?",
+    )
+
+    assert result == "a robot arm"
diff --git a/inference_models/tests/unit_tests/models/test_cosmos3_world.py b/inference_models/tests/unit_tests/models/test_cosmos3_world.py
new file mode 100644
index 0000000000..8242aa4fc6
--- /dev/null
+++ b/inference_models/tests/unit_tests/models/test_cosmos3_world.py
@@ -0,0 +1,115 @@
+from unittest.mock import MagicMock
+
+import numpy as np
+import pytest
+import torch
+
+from inference_models.errors import ModelInputError
+from inference_models.models.cosmos3.cosmos3_world import (
+    SESSION_KEY,
+    Cosmos3ActionTrajectory,
+    Cosmos3EdgeWorldModel,
+)
+
+
+def _world_model() -> Cosmos3EdgeWorldModel:
+    return Cosmos3EdgeWorldModel(runtime=MagicMock(), device=torch.device("cpu"))
+
+
+def _rgb_frames(n: int) -> list:
+    return [np.full((4, 6, 3), 100, dtype=np.uint8) for _ in range(n)]
+
+
+def test_generate_video_converts_output_to_bgr_and_reports_resolution() -> None:
+    model = _world_model()
+    rgb = np.zeros((4, 6, 3), dtype=np.uint8)
+    rgb[..., 0] = 255  # pure red in RGB
+    model._runtime.generate_video.return_value = [rgb, rgb]
+
+    rollout = model.generate_video(
+        images=np.zeros((4, 6, 3), dtype=np.uint8), num_frames=2, fps=24
+    )
+
+    assert len(rollout.frames) == 2
+    assert rollout.frames[0][0, 0].tolist() == [0, 0, 255]  # red lands in BGR[2]
+    assert rollout.resolution == (6, 4)
+    assert rollout.fps == 24.0
+    assert rollout.state_dict is None
+
+
+def test_generate_video_converts_bgr_input_to_rgb_for_runtime() -> None:
+    model = _world_model()
+    model._runtime.generate_video.return_value = []
+    bgr = np.zeros((4, 6, 3), dtype=np.uint8)
+    bgr[..., 0] = 255  # pure blue in BGR
+
+    model.generate_video(images=bgr)
+
+    sent = model._runtime.generate_video.call_args.kwargs["images"][0]
+    assert sent[0, 0].tolist() == [0, 0, 255]  # blue lands in RGB[2]
+
+
+def test_forward_dynamics_without_session_raises() -> None:
+    model = _world_model()
+    trajectory = Cosmos3ActionTrajectory(
+        actions=np.zeros((5, 7)), action_space="droid_joint_velocity", fps=15.0
+    )
+
+    with pytest.raises(ModelInputError):
+        model.forward_dynamics(actions=trajectory, state_dict={})
+    with pytest.raises(ModelInputError):
+        model.forward_dynamics(actions=trajectory, state_dict=None)
+
+
+def test_forward_dynamics_threads_session_state() -> None:
+    model = _world_model()
+    model._runtime.encode_context.return_value = "session-0"
+    model._runtime.rollout.return_value = (_rgb_frames(5), "session-1")
+    trajectory = Cosmos3ActionTrajectory(
+        actions=np.zeros((5, 7)), action_space="droid_joint_velocity", fps=15.0
+    )
+
+    state = model.start_rollout(frames=_rgb_frames(2))
+    rollout = model.forward_dynamics(actions=trajectory, state_dict=state)
+
+    assert state == {SESSION_KEY: "session-0"}
+    assert model._runtime.rollout.call_args.kwargs["session"] == "session-0"
+    assert model._runtime.rollout.call_args.kwargs["num_frames"] == 5
+    assert (
+        model._runtime.rollout.call_args.kwargs["action_space"]
+        == "droid_joint_velocity"
+    )
+    assert rollout.state_dict == {SESSION_KEY: "session-1"}
+    assert len(rollout.frames) == 5
+    assert rollout.fps == 15.0
+
+
+def test_start_rollout_requires_frames() -> None:
+    model = _world_model()
+
+    with pytest.raises(ModelInputError):
+        model.start_rollout(frames=[])
+
+
+def test_inverse_dynamics_wraps_runtime_result() -> None:
+    model = _world_model()
+    model._runtime.infer_actions.return_value = {
+        "actions": [[0.1] * 7] * 3,
+        "action_space": "droid_joint_velocity",
+        "fps": 15,
+        "metadata": {"source": "test"},
+    }
+
+    trajectory = model.inverse_dynamics(
+        frames=_rgb_frames(3), action_space="droid_joint_velocity"
+    )
+
+    assert (
+        model._runtime.infer_actions.call_args.kwargs["action_space"]
+        == "droid_joint_velocity"
+    )
+
+    assert trajectory.actions.shape == (3, 7)
+    assert trajectory.action_space == "droid_joint_velocity"
+    assert trajectory.fps == 15.0
+    assert trajectory.metadata == {"source": "test"}
diff --git a/inference_models/uv.lock b/inference_models/uv.lock
index 3a27640102..77fc081d77 100644
--- a/inference_models/uv.lock
+++ b/inference_models/uv.lock
@@ -1,5 +1,5 @@
 version = 1
-revision = 3
+revision = 2
 requires-python = ">=3.10, <3.13"
 resolution-markers = [
     "python_full_version >= '3.12' and platform_machine != 's390x' and sys_platform == 'darwin'",
@@ -139,7 +139,7 @@ name = "anyio"
 version = "4.12.1"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
-    { name = "exceptiongroup", marker = "python_full_version < '3.11' or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu118') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu12' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu118') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu124') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu124') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu130' and extra == 'extra-16-inference-models-torch-jp6-cu126')" },
+    { name = "exceptiongroup", marker = "python_full_version < '3.11' or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu118') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu12' and extra == 'extra-16-inference-models-onnx-jp6-cu126')" },
     { name = "idna" },
     { name = "typing-extensions" },
 ]
@@ -390,7 +390,7 @@ name = "click"
 version = "8.3.1"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
-    { name = "colorama", marker = "sys_platform == 'win32' or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu118') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu12' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu118') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu124') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu124') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu130' and extra == 'extra-16-inference-models-torch-jp6-cu126')" },
+    { name = "colorama", marker = "sys_platform == 'win32' or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu118') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu12' and extra == 'extra-16-inference-models-onnx-jp6-cu126')" },
 ]
 sdist = { url = "https://files.pythonhosted.org/packages/3d/fa/656b739db8587d7b5dfa22e22ed02566950fbfbcdc20311993483657a5c0/click-8.3.1.tar.gz", hash = "sha256:12ff4785d337a1bb490bb7e9c2b1ee5da3112e94a8622f26a6c77f5d2fc6842a", size = 295065, upload-time = "2025-11-15T20:45:42.706Z" }
 wheels = [
@@ -411,7 +411,7 @@ name = "coloredlogs"
 version = "15.0.1"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
-    { name = "humanfriendly", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu118') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu124') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu126') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu128') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu130') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu124') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu126') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu128') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu130') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu126') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu128') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu130') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu128') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu130') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-cu130') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cu130' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu118') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu124') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu126') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu128') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu130') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu124') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu126') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu128') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu130') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu126') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu128') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu130') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu128') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu130') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-cu130') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cu130' and extra == 'extra-16-inference-models-torch-jp6-cu126') or extra == 'extra-16-inference-models-onnx-cpu' or extra == 'extra-16-inference-models-onnx-cu118' or extra == 'extra-16-inference-models-onnx-cu12' or (extra != 'extra-16-inference-models-onnx-jp6-cu126' and extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu118') or (extra != 'extra-16-inference-models-onnx-jp6-cu126' and extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu124') or (extra != 'extra-16-inference-models-onnx-jp6-cu126' and extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu126') or (extra != 'extra-16-inference-models-onnx-jp6-cu126' and extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu128') or (extra != 'extra-16-inference-models-onnx-jp6-cu126' and extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu130') or (extra != 'extra-16-inference-models-onnx-jp6-cu126' and extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra != 'extra-16-inference-models-onnx-jp6-cu126' and extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu124') or (extra != 'extra-16-inference-models-onnx-jp6-cu126' and extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu126') or (extra != 'extra-16-inference-models-onnx-jp6-cu126' and extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu128') or (extra != 'extra-16-inference-models-onnx-jp6-cu126' and extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu130') or (extra != 'extra-16-inference-models-onnx-jp6-cu126' and extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra != 'extra-16-inference-models-onnx-jp6-cu126' and extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu126') or (extra != 'extra-16-inference-models-onnx-jp6-cu126' and extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu128') or (extra != 'extra-16-inference-models-onnx-jp6-cu126' and extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu130') or (extra != 'extra-16-inference-models-onnx-jp6-cu126' and extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra != 'extra-16-inference-models-onnx-jp6-cu126' and extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu128') or (extra != 'extra-16-inference-models-onnx-jp6-cu126' and extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu130') or (extra != 'extra-16-inference-models-onnx-jp6-cu126' and extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra != 'extra-16-inference-models-onnx-jp6-cu126' and extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-cu130') or (extra != 'extra-16-inference-models-onnx-jp6-cu126' and extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra != 'extra-16-inference-models-onnx-jp6-cu126' and extra == 'extra-16-inference-models-torch-cu130' and extra == 'extra-16-inference-models-torch-jp6-cu126')" },
+    { name = "humanfriendly", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu118') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu124') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu126') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu128') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu130') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu124') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu126') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu128') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu130') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu126') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu128') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu130') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu128') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu130') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-cu130') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (platform_machine != 'aarch64' and extra == 'extra-16-inference-models-torch-cu130' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu118') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu124') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu126') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu128') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu130') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu124') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu126') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu128') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu130') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu126') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu128') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu130') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu128') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu130') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-cu130') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (sys_platform != 'linux' and extra == 'extra-16-inference-models-torch-cu130' and extra == 'extra-16-inference-models-torch-jp6-cu126')" },
 ]
 sdist = { url = "https://files.pythonhosted.org/packages/cc/c7/eed8f27100517e8c0e6b923d5f0845d0cb99763da6fdee00478f91db7325/coloredlogs-15.0.1.tar.gz", hash = "sha256:7c991aa71a4577af2f82600d8f8f3a89f936baeaf9b50a9c197da014e5bf16b0", size = 278520, upload-time = "2021-06-11T10:22:45.202Z" }
 wheels = [
@@ -590,7 +590,7 @@ name = "exceptiongroup"
 version = "1.3.0"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
-    { name = "typing-extensions", marker = "python_full_version < '3.11' or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu118') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu12' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu118') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu124') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu124') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu130' and extra == 'extra-16-inference-models-torch-jp6-cu126')" },
+    { name = "typing-extensions", marker = "python_full_version < '3.11' or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu118') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu12' and extra == 'extra-16-inference-models-onnx-jp6-cu126')" },
 ]
 sdist = { url = "https://files.pythonhosted.org/packages/0b/9f/a65090624ecf468cdca03533906e7c69ed7588582240cfe7cc9e770b50eb/exceptiongroup-1.3.0.tar.gz", hash = "sha256:b241f5885f560bc56a59ee63ca4c6a8bfa46ae4ad651af316d4e81817bb9fd88", size = 29749, upload-time = "2025-05-10T17:42:51.123Z" }
 wheels = [
@@ -829,7 +829,7 @@ source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "filelock" },
     { name = "fsspec" },
-    { name = "hf-xet", marker = "platform_machine == 'AMD64' or platform_machine == 'aarch64' or platform_machine == 'amd64' or platform_machine == 'arm64' or platform_machine == 'x86_64' or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu118') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu12' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu118') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu124') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu124') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu130' and extra == 'extra-16-inference-models-torch-jp6-cu126')" },
+    { name = "hf-xet", marker = "platform_machine == 'AMD64' or platform_machine == 'aarch64' or platform_machine == 'amd64' or platform_machine == 'arm64' or platform_machine == 'x86_64' or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu118') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu12' and extra == 'extra-16-inference-models-onnx-jp6-cu126')" },
     { name = "httpx" },
     { name = "packaging" },
     { name = "pyyaml" },
@@ -1896,7 +1896,7 @@ resolution-markers = [
     "(python_full_version < '3.11' and platform_machine != 'aarch64' and platform_machine != 's390x' and sys_platform == 'linux') or (python_full_version < '3.11' and platform_machine != 's390x' and sys_platform != 'darwin' and sys_platform != 'linux')",
 ]
 dependencies = [
-    { name = "numpy", marker = "platform_machine != 's390x' or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu118') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu12' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu118') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu124') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu124') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu130' and extra == 'extra-16-inference-models-torch-jp6-cu126')" },
+    { name = "numpy", marker = "platform_machine != 's390x' or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu118') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu124') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu124') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu130' and extra == 'extra-16-inference-models-torch-jp6-cu126')" },
 ]
 sdist = { url = "https://files.pythonhosted.org/packages/32/49/6e67c334872d2c114df3020e579f3718c333198f8312290e09ec0216703a/ml_dtypes-0.5.1.tar.gz", hash = "sha256:ac5b58559bb84a95848ed6984eb8013249f90b6bab62aa5acbad876e256002c9", size = 698772, upload-time = "2025-01-07T03:34:55.613Z" }
 wheels = [
@@ -1927,7 +1927,7 @@ resolution-markers = [
     "python_full_version < '3.11' and platform_machine == 's390x' and sys_platform != 'darwin'",
 ]
 dependencies = [
-    { name = "numpy", marker = "platform_machine == 's390x' or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu118') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu12' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu118') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu124') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu124') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu130' and extra == 'extra-16-inference-models-torch-jp6-cu126')" },
+    { name = "numpy", marker = "platform_machine == 's390x' or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu118') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu124') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu124') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu130' and extra == 'extra-16-inference-models-torch-jp6-cu126')" },
 ]
 sdist = { url = "https://files.pythonhosted.org/packages/0e/4a/c27b42ed9b1c7d13d9ba8b6905dece787d6259152f2309338aed29b2447b/ml_dtypes-0.5.4.tar.gz", hash = "sha256:8ab06a50fb9bf9666dd0fe5dfb4676fa2b0ac0f31ecff72a6c3af8e22c063453", size = 692314, upload-time = "2025-11-17T22:32:31.031Z" }
 wheels = [
@@ -3032,6 +3032,15 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/bc/60/5382c03e1970de634027cee8e1b7d39776b778b81812aaf45b694dfe9e28/pillow-12.2.0-pp311-pypy311_pp73-win_amd64.whl", hash = "sha256:bfa9c230d2fe991bed5318a5f119bd6780cda2915cca595393649fc118ab895e", size = 7080946, upload-time = "2026-04-01T14:46:11.734Z" },
 ]
 
+[[package]]
+name = "pkgconfig"
+version = "1.6.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/52/fd/0adde075cd3bfecd557bc7d757e00e231d34d8a6edb4c8d1642759254c21/pkgconfig-1.6.0.tar.gz", hash = "sha256:4a5a6631ce937fafac457104a40d558785a658bbdca5c49b6295bc3fd651907f", size = 5691, upload-time = "2026-03-06T11:26:01.194Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/75/6f/f7ec07fba48f07c555cc4099481df644fbbc12067879072c17ac229f6556/pkgconfig-1.6.0-py3-none-any.whl", hash = "sha256:98e71754855e9563838d952a160eb577edabb57782e49853edb5381927e6bea1", size = 7086, upload-time = "2026-03-06T11:26:07.688Z" },
+]
+
 [[package]]
 name = "platformdirs"
 version = "4.3.8"
@@ -3055,7 +3064,7 @@ name = "portalocker"
 version = "3.2.0"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
-    { name = "pywin32", marker = "sys_platform == 'win32' or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu118') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu12' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu118') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu124') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu124') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu130' and extra == 'extra-16-inference-models-torch-jp6-cu126')" },
+    { name = "pywin32", marker = "sys_platform == 'win32' or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu118') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu12' and extra == 'extra-16-inference-models-onnx-jp6-cu126')" },
 ]
 sdist = { url = "https://files.pythonhosted.org/packages/5e/77/65b857a69ed876e1951e88aaba60f5ce6120c33703f7cb61a3c894b8c1b6/portalocker-3.2.0.tar.gz", hash = "sha256:1f3002956a54a8c3730586c5c77bf18fae4149e07eaf1c29fc3faf4d5a3f89ac", size = 95644, upload-time = "2025-06-14T13:20:40.03Z" }
 wheels = [
@@ -3539,6 +3548,7 @@ version = "2.2.3"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "cffi" },
+    { name = "pkgconfig" },
 ]
 sdist = { url = "https://files.pythonhosted.org/packages/7b/88/f73dae807ec68b228fba72507105e3ba80a561dc0bade0004ce24fd118fc/pyvips-2.2.3.tar.gz", hash = "sha256:43bceced0db492654c93008246a58a508e0373ae1621116b87b322f2ac72212f", size = 56626, upload-time = "2024-04-28T11:19:58.158Z" }
 
@@ -4560,9 +4570,9 @@ resolution-markers = [
     "python_full_version < '3.11' and platform_machine == 's390x' and sys_platform != 'darwin'",
 ]
 dependencies = [
-    { name = "huggingface-hub", marker = "sys_platform != 'darwin' or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu118') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu12' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu118') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu124') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu124') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu130' and extra == 'extra-16-inference-models-torch-jp6-cu126')" },
-    { name = "pyyaml", marker = "sys_platform != 'darwin' or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu118') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu12' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu118') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu124') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu124') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu130' and extra == 'extra-16-inference-models-torch-jp6-cu126')" },
-    { name = "safetensors", marker = "sys_platform != 'darwin' or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu118') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu12' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu118') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu124') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu124') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu130' and extra == 'extra-16-inference-models-torch-jp6-cu126')" },
+    { name = "huggingface-hub", marker = "sys_platform != 'darwin' or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu118') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu12' and extra == 'extra-16-inference-models-onnx-jp6-cu126')" },
+    { name = "pyyaml", marker = "sys_platform != 'darwin' or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu118') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu12' and extra == 'extra-16-inference-models-onnx-jp6-cu126')" },
+    { name = "safetensors", marker = "sys_platform != 'darwin' or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu118') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu12' and extra == 'extra-16-inference-models-onnx-jp6-cu126')" },
     { name = "torch", version = "2.6.0+cu124", source = { registry = "https://download.pytorch.org/whl/cu124" }, marker = "(sys_platform != 'darwin' and extra != 'extra-16-inference-models-onnx-jp6-cu126' and extra != 'extra-16-inference-models-torch-cpu' and extra != 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu124') or (sys_platform != 'darwin' and extra != 'extra-16-inference-models-onnx-cpu' and extra != 'extra-16-inference-models-onnx-cu118' and extra != 'extra-16-inference-models-onnx-cu12' and extra == 'extra-16-inference-models-torch-cu124') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu118') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu12' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu118') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu124') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu124') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu130' and extra == 'extra-16-inference-models-torch-jp6-cu126')" },
     { name = "torch", version = "2.7.1+cpu", source = { registry = "https://download.pytorch.org/whl/cpu" }, marker = "(sys_platform != 'darwin' and extra != 'extra-16-inference-models-onnx-jp6-cu126' and extra == 'extra-16-inference-models-torch-cpu') or (sys_platform != 'darwin' and extra != 'extra-16-inference-models-onnx-cpu' and extra != 'extra-16-inference-models-onnx-cu118' and extra != 'extra-16-inference-models-onnx-cu12' and extra == 'extra-16-inference-models-torch-cpu') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu118') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu12' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu118') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu124') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu124') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu130' and extra == 'extra-16-inference-models-torch-jp6-cu126')" },
     { name = "torch", version = "2.7.1+cu118", source = { registry = "https://download.pytorch.org/whl/cu118" }, marker = "(sys_platform != 'darwin' and extra != 'extra-16-inference-models-onnx-jp6-cu126' and extra != 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu118') or (sys_platform != 'darwin' and extra != 'extra-16-inference-models-onnx-cpu' and extra != 'extra-16-inference-models-onnx-cu118' and extra != 'extra-16-inference-models-onnx-cu12' and extra == 'extra-16-inference-models-torch-cu118') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu118') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu12' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu118') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu124') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu124') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu130' and extra == 'extra-16-inference-models-torch-jp6-cu126')" },
@@ -5251,7 +5261,7 @@ name = "tqdm"
 version = "4.67.1"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
-    { name = "colorama", marker = "sys_platform == 'win32' or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu118') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu12' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu118') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu124') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cpu' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu124') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu118' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu126') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu124' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu128') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu126' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-cu130') or (extra == 'extra-16-inference-models-torch-cu128' and extra == 'extra-16-inference-models-torch-jp6-cu126') or (extra == 'extra-16-inference-models-torch-cu130' and extra == 'extra-16-inference-models-torch-jp6-cu126')" },
+    { name = "colorama", marker = "sys_platform == 'win32' or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu118') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cpu' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-cu12') or (extra == 'extra-16-inference-models-onnx-cu118' and extra == 'extra-16-inference-models-onnx-jp6-cu126') or (extra == 'extra-16-inference-models-onnx-cu12' and extra == 'extra-16-inference-models-onnx-jp6-cu126')" },
 ]
 sdist = { url = "https://files.pythonhosted.org/packages/a8/4b/29b4ef32e036bb34e4ab51796dd745cdba7ed47ad142a9f4a1eb8e0c744d/tqdm-4.67.1.tar.gz", hash = "sha256:f8aef9c52c08c13a65f30ea34f4e5aac3fd1a34959879d7e59e63027286627f2", size = 169737, upload-time = "2024-11-24T20:12:22.481Z" }
 wheels = [
diff --git a/requirements/requirements.cosmos.txt b/requirements/requirements.cosmos.txt
new file mode 100644
index 0000000000..51164a4116
--- /dev/null
+++ b/requirements/requirements.cosmos.txt
@@ -0,0 +1,13 @@
+# Optional dependencies for NVIDIA Cosmos 3 Edge (reasoner + generator).
+#
+# Cosmos 3 support is newer than any transformers/diffusers release:
+# - transformers: `cosmos3_edge` model type requires >=5.15 (unreleased)
+# - diffusers: Cosmos3OmniPipeline checkpoint format requires >=0.40 (unreleased;
+#   0.39.0 loads the transformer with meta/randomly-initialized tensors)
+#
+# Until those releases land we pin the exact git commits validated on real
+# weights (L4, 2026-07-20). Replace with pypi version pins once released --
+# git direct-URL deps cannot ship in the pypi extras (see .release/pypi/*.setup.py),
+# so this file is intentionally NOT wired into extras_require yet.
+transformers @ git+https://github.com/huggingface/transformers.git@cbf4d720ec734edb77d25452b2790c7e4be2f8d7
+diffusers @ git+https://github.com/huggingface/diffusers.git@86e6dac5360703ddf09fe250db50be667eb93662
diff --git a/requirements/requirements.cpu.txt b/requirements/requirements.cpu.txt
index 78a211734c..2fe3ce4eaf 100644
--- a/requirements/requirements.cpu.txt
+++ b/requirements/requirements.cpu.txt
@@ -1,3 +1,3 @@
 onnxruntime>=1.15.1,<1.22.0
 nvidia-ml-py<13.0.0
-inference-models[torch-cpu,onnx-cpu]~=0.32.0rc3  # keep in sync between requirements.gpu.txt, requirements.cpu.txt, requirements.vino.txt
+inference-models[torch-cpu,onnx-cpu]~=0.32.0  # keep in sync between requirements.gpu.txt, requirements.cpu.txt, requirements.vino.txt
diff --git a/requirements/requirements.gpu.txt b/requirements/requirements.gpu.txt
index 6e5bd7c61b..4f89ab253b 100644
--- a/requirements/requirements.gpu.txt
+++ b/requirements/requirements.gpu.txt
@@ -1,2 +1,2 @@
 onnxruntime-gpu>=1.15.1,<1.22.0
-inference-models[torch-cu124,onnx-cu12]~=0.32.0rc3 # keep in sync between requirements.jetson requirements.gpu.txt, requirements.cpu.txt, requirements.vino.txt
+inference-models[torch-cu124,onnx-cu12]~=0.32.0 # keep in sync between requirements.jetson requirements.gpu.txt, requirements.cpu.txt, requirements.vino.txt
diff --git a/requirements/requirements.jetson.txt b/requirements/requirements.jetson.txt
index f94a3d7e03..5c4a0f96ca 100644
--- a/requirements/requirements.jetson.txt
+++ b/requirements/requirements.jetson.txt
@@ -1,3 +1,3 @@
 pypdfium2>=4.11.0,<5.0.0
 PyYAML~=6.0.0
-inference-models~=0.32.0rc3  # keep in sync between requirements.jetson requirements.gpu.txt, requirements.cpu.txt, requirements.vino.txt
+inference-models~=0.32.0  # keep in sync between requirements.jetson requirements.gpu.txt, requirements.cpu.txt, requirements.vino.txt
diff --git a/requirements/requirements.vino.txt b/requirements/requirements.vino.txt
index fc2e6bcdf6..190ca0127e 100644
--- a/requirements/requirements.vino.txt
+++ b/requirements/requirements.vino.txt
@@ -1,2 +1,2 @@
 onnxruntime-openvino>=1.15.0,<1.22.0
-inference-models[torch-cpu]~=0.32.0rc3  # keep in sync between requirements.jetson requirements.gpu.txt, requirements.cpu.txt, requirements.vino.txt
+inference-models[torch-cpu]~=0.32.0  # keep in sync between requirements.jetson requirements.gpu.txt, requirements.cpu.txt, requirements.vino.txt
diff --git a/tests/workflows/unit_tests/core_steps/models/foundation/test_cosmos3.py b/tests/workflows/unit_tests/core_steps/models/foundation/test_cosmos3.py
new file mode 100644
index 0000000000..1a8e5ac4e9
--- /dev/null
+++ b/tests/workflows/unit_tests/core_steps/models/foundation/test_cosmos3.py
@@ -0,0 +1,71 @@
+"""Unit tests for the Cosmos 3 Edge v1 block manifest."""
+
+import pytest
+from pydantic import ValidationError
+
+from inference.core.workflows.core_steps.models.foundation.cosmos3.v1 import (
+    BlockManifest,
+    _combine_prompt,
+)
+
+BASE = {
+    "type": "roboflow_core/cosmos3_edge@v1",
+    "name": "my_cosmos_step",
+    "images": "$inputs.image",
+}
+
+
+def test_manifest_parses_with_defaults():
+    result = BlockManifest.model_validate(BASE)
+    assert result.prompt is None
+    assert result.system_prompt is None
+    assert result.model_version == "nvidia/cosmos-3-edge"
+
+
+def test_manifest_accepts_literal_prompts():
+    result = BlockManifest.model_validate(
+        {**BASE, "prompt": "Is the path clear?", "system_prompt": "Be terse."}
+    )
+    assert result.prompt == "Is the path clear?"
+    assert result.system_prompt == "Be terse."
+
+
+def test_manifest_accepts_input_selector_prompts():
+    result = BlockManifest.model_validate(
+        {
+            **BASE,
+            "prompt": "$inputs.prompt",
+            "system_prompt": "$inputs.system_prompt",
+        }
+    )
+    assert result.prompt == "$inputs.prompt"
+    assert result.system_prompt == "$inputs.system_prompt"
+
+
+def test_manifest_accepts_model_version_selector():
+    result = BlockManifest.model_validate(
+        {**BASE, "model_version": "$inputs.model_version"}
+    )
+    assert result.model_version == "$inputs.model_version"
+
+
+def test_manifest_rejects_invalid_image_selector():
+    with pytest.raises(ValidationError):
+        BlockManifest.model_validate({**BASE, "images": 42})
+
+
+def test_manifest_declares_single_output():
+    outputs = BlockManifest.describe_outputs()
+    assert [o.name for o in outputs] == ["output"]
+
+
+def test_combine_prompt_uses_sentinel():
+    assert (
+        _combine_prompt(prompt="Question?", system_prompt="Context.")
+        == "Question?<system_prompt>Context."
+    )
+
+
+def test_combine_prompt_applies_defaults():
+    combined = _combine_prompt(prompt=None, system_prompt=None)
+    assert combined.startswith("Describe what's in this image.<system_prompt>")