Describe the bug
MLXLLM does not recognize the Gemma 4 model architecture. When attempting to load any Gemma 4 MLX model (e.g., mlx-community/gemma-4-E2B-it-4bit), the model fails to load because the gemma4 architecture type is not registered in the model registry. Gemma 4 was released on April 2, 2026 with MLX-converted weights available on Hugging Face from day one, but the Swift library doesn't support loading them.
To Reproduce
import MLXLLM
import MLXLMCommon
let modelConfig = ModelConfiguration(id: "mlx-community/gemma-4-E2B-it-4bit")
let container = try await LLMModelFactory.shared.loadContainer(configuration: modelConfig)
// Fails: model type "gemma4" is not supported
Expected behavior
MLXLLM should recognize and load Gemma 4 models, including the new architectural features:
- Per-Layer Embeddings (PLE)
- Shared KV Cache (later layers reuse KV states from earlier layers)
- Dual RoPE (standard RoPE for sliding-window layers, proportional RoPE for global layers)
- MoE variant (26B model with 3.8B active parameters)
The Python mlx_lm and mlx-vlm libraries added Gemma 4 support on launch day (v0.4.3), so the model weights and conversion are already validated.
Desktop (please complete the following information):
- OS Version: macOS 26.3.1
- Device: M2 MacBook Pro, 24 GB
- Version: mlx-swift 0.10.0, mlx-swift-examples 2.29.1
Additional context
Available MLX-converted Gemma 4 models on Hugging Face:
mlx-community/gemma-4-26b-a4b-it-4bit (MoE, 3.8B active)mlx-community/gemma-4-31b-it-4bit (Dense)unsloth/gemma-4-E2B-it-UD-MLX-4bit (Edge, ~2B)unsloth/gemma-4-E4B-it-UD-MLX-4bit (Edge, ~4B)
Full collection: https://huggingface.co/collections/mlx-community/gemma-4
The Python MLX ecosystem already supports these models — this is a request to bring parity to the Swift side.
Describe the bug
MLXLLM does not recognize the Gemma 4 model architecture. When attempting to load any Gemma 4 MLX model (e.g.,
mlx-community/gemma-4-E2B-it-4bit), the model fails to load because thegemma4architecture type is not registered in the model registry. Gemma 4 was released on April 2, 2026 with MLX-converted weights available on Hugging Face from day one, but the Swift library doesn't support loading them.To Reproduce
Expected behavior
MLXLLM should recognize and load Gemma 4 models, including the new architectural features:
The Python
mlx_lmandmlx-vlmlibraries added Gemma 4 support on launch day (v0.4.3), so the model weights and conversion are already validated.Desktop (please complete the following information):
Additional context
Available MLX-converted Gemma 4 models on Hugging Face:
mlx-community/gemma-4-26b-a4b-it-4bit(MoE, 3.8B active)mlx-community/gemma-4-31b-it-4bit(Dense)unsloth/gemma-4-E2B-it-UD-MLX-4bit(Edge, ~2B)unsloth/gemma-4-E4B-it-UD-MLX-4bit(Edge, ~4B)Full collection: https://huggingface.co/collections/mlx-community/gemma-4
The Python MLX ecosystem already supports these models — this is a request to bring parity to the Swift side.