ai: qwen3.5:2b is the global MiOS chat default

Operator directive 2026-05-11: 'qwen3.5:2b make it the GLOBAL MiOS
defaults'. Flip [ai].model and the auto-pick small/mid/big chain
from the qwen2.5-coder family to qwen3:

  [ai].model                       qwen2.5-coder:7b  -> qwen3.5:2b
  [ai].bake_models                 qwen2.5-coder:7b  -> qwen3.5:2b
  [ai.host_thresholds].small       phi4-mini:3.8b... -> qwen3.5:2b
  [ai.host_thresholds].mid         qwen2.5-coder:7b  -> qwen3.5:7b
  [ai.host_thresholds].big         qwen2.5-coder:14b -> qwen3.5:14b

Mirror the same flip into every SSOT consumer that hard-references
the model:

  usr/share/mios/profile.toml         (layered above mios.toml)
  usr/libexec/mios/ollama-firstboot.sh  DEFAULT_CHAT
  usr/share/mios/hermes/config.yaml    backend.model + auxiliary.model
                                       + available_models
  etc/containers/systemd/mios-hermes.container  HERMES_BACKEND_MODEL
  usr/share/mios/distrobox/aichat/config.yaml  default model alias
  etc/aichat/config.yaml               default model alias
  usr/bin/mios                         MIOS_AI_MODEL default
  usr/bin/mios-ai                      MIOS_AI_MODEL default

Operator-pickable from mios.html; overrideable per-install via the
MIOS_AI_MODEL env var. Docs / examples / training data still cite
qwen2.5-coder:7b -- those trail the SSOT and update on next sweep.

Author: Kabuki94

etc/aichat/config.yaml

@@ -14,15 +14,15 @@
 # default chat + embed models are present and aichat can connect
 # without any user-side config.
 #
-# Model: pinned to qwen2.5-coder:7b as a safe default. Override per
+# Model: pinned to qwen3.5:2b as a safe default. Override per
 # session: `aichat --model ollama:<model>`. Persistent override:
 # copy this file to ~/.config/aichat/config.yaml and edit `model`.
 # The MIOS_AI_MODEL env var (set by /etc/profile.d/mios-env.sh from
 # /etc/mios/install.env) is the canonical source-of-truth; the
 # /usr/bin/mios-ai wrapper substitutes it at call time so the bare
 # `aichat` and `mios-ai` paths can drift apart cleanly.
 
-model: ollama:qwen2.5-coder:7b
+model: ollama:qwen3.5:2b
 stream: true
 save_session: true
 keybindings: emacs
@@ -33,7 +33,7 @@ clients:
     api_base: http://localhost:11434/v1
     api_key: ollama   # any non-empty string; ollama ignores the key
     models:
-      - name: qwen2.5-coder:7b
+      - name: qwen3.5:2b
         max_input_tokens: 32768
         supports_function_calling: true
       - name: qwen2.5-coder:14b
@@ -52,5 +52,5 @@ clients:
     api_base: http://localhost:8080/v1
     api_key: localai
     models:
-      - name: qwen2.5-coder:7b
+      - name: qwen3.5:2b
         max_input_tokens: 32768

etc/containers/systemd/mios-hermes.container

@@ -57,7 +57,7 @@ EnvironmentFile=/etc/mios/hermes/api.env
 # interactively; we bake them in so the unit is fully declarative.
 Environment=HERMES_BACKEND_PROVIDER=ollama
 Environment=HERMES_BACKEND_BASE_URL=http://mios-ollama:11434
-Environment=HERMES_BACKEND_MODEL=qwen2.5-coder:7b
+Environment=HERMES_BACKEND_MODEL=qwen3.5:2b
 Environment=HERMES_CONFIG_PATH=/etc/hermes/config.yaml
 
 # Identity. UID resolved from mios.toml [services.hermes].uid (default

usr/bin/mios

@@ -17,7 +17,7 @@
 # Environment (per Architectural Law 5):
 #   MIOS_AI_ENDPOINT=http://localhost:8080/v1   the OpenAI-compatible base URL
 #   MIOS_AI_KEY=                                bearer token (empty OK locally)
-#   MIOS_AI_MODEL=qwen2.5-coder:7b              the chat model id
+#   MIOS_AI_MODEL=qwen3.5:2b              the chat model id
 #   MIOS_AI_SYSTEM_PROMPT=/path/to/system.md    explicit override
 #   MIOS_SHARE_DIR=/usr/share/mios              base for resolved files
 #
@@ -47,7 +47,7 @@ except ImportError:
 
 # ─── Configuration ────────────────────────────────────────────────────
 ENDPOINT = os.environ.get("MIOS_AI_ENDPOINT", "http://localhost:8080/v1")
-MODEL    = os.environ.get("MIOS_AI_MODEL",    "qwen2.5-coder:7b")
+MODEL    = os.environ.get("MIOS_AI_MODEL",    "qwen3.5:2b")
 API_KEY  = os.environ.get("MIOS_AI_KEY",      "sk-noop")
 SHARE    = pathlib.Path(os.environ.get("MIOS_SHARE_DIR", "/usr/share/mios"))
 LIB      = pathlib.Path("/usr/lib/mios")

usr/bin/mios-ai

@@ -16,7 +16,7 @@
 #   1. CLI args (highest)
 #   2. $MIOS_AI_ENDPOINT / $MIOS_AI_MODEL (env, set by mios-env.sh)
 #   3. /etc/mios/install.env defaults
-#   4. Hardcoded fallbacks (mios @ qwen2.5-coder:7b)
+#   4. Hardcoded fallbacks (mios @ qwen3.5:2b)
 #
 # Routes through aichat. /usr/bin/aichat on a deployed MiOS host is the
 # distrobox-routed shim (see automation/37-aichat.sh + the
@@ -63,9 +63,9 @@ if [[ -r /etc/mios/install.env ]]; then
     set -a; source /etc/mios/install.env 2>/dev/null || true; set +a
 fi
 
-MIOS_AI_MODEL="${MODEL_OVERRIDE:-${MIOS_AI_MODEL:-qwen2.5-coder:7b}}"
+MIOS_AI_MODEL="${MODEL_OVERRIDE:-${MIOS_AI_MODEL:-qwen3.5:2b}}"
 # Strip any client-prefix that may have been set in install.env or env
-# (e.g. MIOS_AI_MODEL=mios:qwen2.5-coder:7b). The CLIENT name is
+# (e.g. MIOS_AI_MODEL=mios:qwen3.5:2b). The CLIENT name is
 # composed below; this leaves us with just the model.
 case "$MIOS_AI_MODEL" in
     mios:*|ollama:*|localai:*) MIOS_AI_MODEL="${MIOS_AI_MODEL#*:}" ;;

usr/libexec/mios/ollama-firstboot.sh

@@ -64,7 +64,7 @@ fi
 # ── Layer 2: ensure the configured models are present, pulling any gaps ──
 # Read the runtime model selection from install.env so operators can
 # override the build-baked default without rebuilding the image.
-DEFAULT_CHAT="qwen2.5-coder:7b"
+DEFAULT_CHAT="qwen3.5:2b"
 DEFAULT_EMBED="nomic-embed-text"
 MIOS_AI_MODEL="$DEFAULT_CHAT"
 MIOS_AI_EMBED_MODEL="$DEFAULT_EMBED"

usr/share/mios/distrobox/aichat/config.yaml

@@ -13,9 +13,9 @@
 #
 # `ollama` and `localai` remain available as named alternates for
 # operators who want to switch backends without editing the default
-# (e.g. `aichat --model ollama:qwen2.5-coder:7b "..."`).
+# (e.g. `aichat --model ollama:qwen3.5:2b "..."`).
 
-model: mios:qwen2.5-coder:7b
+model: mios:qwen3.5:2b
 stream: true
 save_session: true
 keybindings: emacs
@@ -26,7 +26,7 @@ clients:
     api_base: http://localhost:8080/v1
     api_key: localai
     models:
-      - name: qwen2.5-coder:7b
+      - name: qwen3.5:2b
         max_input_tokens: 32768
         supports_function_calling: true
       - name: qwen2.5-coder:14b
@@ -41,7 +41,7 @@ clients:
     api_base: http://localhost:11434/v1
     api_key: ollama
     models:
-      - name: qwen2.5-coder:7b
+      - name: qwen3.5:2b
         max_input_tokens: 32768
         supports_function_calling: true
       - name: qwen2.5-coder:14b
@@ -56,5 +56,5 @@ clients:
     api_base: http://localhost:8080/v1
     api_key: localai
     models:
-      - name: qwen2.5-coder:7b
+      - name: qwen3.5:2b
         max_input_tokens: 32768

usr/share/mios/hermes/config.yaml

@@ -10,18 +10,18 @@
 # This file replaces the interactive `setup` wizard that
 # deploy-aistack.sh runs on bare-metal first-deploy. Wiring is fully
 # declarative: backend = mios-ollama on ai-net, model = MiOS's baked
-# default (qwen2.5-coder:7b), API server reads its key from
+# default (qwen3.5:2b), API server reads its key from
 # /etc/mios/hermes/api.env via EnvironmentFile= in the Quadlet.
 
 backend:
   provider: ollama
   base_url: http://mios-ollama:11434
-  model: qwen2.5-coder:7b
+  model: qwen3.5:2b
   # Listed alternates so the WebUI's model picker shows what's
-  # available. mios-ollama-firstboot bakes qwen2.5-coder:7b and
+  # available. mios-ollama-firstboot bakes qwen3.5:2b and
   # nomic-embed-text by default; operators can pull others post-deploy.
   available_models:
-    - qwen2.5-coder:7b
+    - qwen3.5:2b
     - nomic-embed-text
 
 # Auxiliary LLM -- Hermes uses this for context compression,
@@ -37,7 +37,7 @@ backend:
 auxiliary:
   provider: local/custom
   base_url: http://mios-ai:8080/v1
-  model: qwen2.5-coder:7b
+  model: qwen3.5:2b
   api_key: ${API_SERVER_KEY}
 
 api_server:

usr/share/mios/mios.toml

@@ -403,7 +403,10 @@ enable   = true
 # Ollama. Operator-flagged 2026-05-11 (`mios hello` -> 500 from
 # LocalAI before this fix).
 endpoint            = "http://localhost:11434/v1"
-model               = "qwen2.5-coder:7b"
+# Global MiOS chat default. Operator directive 2026-05-11: "qwen3.5:2b
+# make it the GLOBAL MiOS defaults". Operator-pickable from mios.html;
+# overrideable per-install via MIOS_AI_MODEL.
+model               = "qwen3.5:2b"
 embed_model         = "nomic-embed-text"
 api_key             = ""                # empty for localhost; cloud needs a key
 enable_ollama       = true
@@ -413,7 +416,7 @@ system_prompt_file  = ""                # optional override; empty = use canonic
 # Build-time model bake list. Comma-separated; consumed by
 # automation/37-ollama-prep.sh as MIOS_OLLAMA_BAKE_MODELS. Empty disables
 # baking (image stays small; first-boot pulls instead).
-bake_models         = "qwen2.5-coder:7b,nomic-embed-text"
+bake_models         = "qwen3.5:2b,nomic-embed-text"
 # Recommended floor for the default model. Informational; logged in
 # postcheck so operators know what they're signing up for.
 ram_floor_gb        = 12
@@ -431,11 +434,14 @@ runtime_dir         = "/var/lib/ollama/models"        # writable, first-boot-see
 # these and assigns $aiModel = big/mid/small at install time.
 # ----------------------------------------------------------------------------
 [ai.host_thresholds]
+# Auto-pick rolls UP from the small default (qwen3.5:2b at < 12 GB
+# RAM). Operators with 12+ GB get qwen3.5:7b; 32+ GB get qwen3.5:14b.
+# All three are operator-pickable from mios.html.
 big_ram_gb       = 32                            # >= this -> big_ram_model
 mid_ram_gb       = 12                            # >= this -> mid_ram_model
-big_ram_model    = "qwen2.5-coder:14b"
-mid_ram_model    = "qwen2.5-coder:7b"
-small_ram_model  = "phi4-mini:3.8b-q4_K_M"      # < mid_ram_gb fallback
+big_ram_model    = "qwen3.5:14b"
+mid_ram_model    = "qwen3.5:7b"
+small_ram_model  = "qwen3.5:2b"                  # < mid_ram_gb fallback (also the global default)
 
 # ----------------------------------------------------------------------------
 # [desktop] -- session preferences and Flatpak picks.

usr/share/mios/profile.toml

@@ -46,8 +46,9 @@ allow_cockpit          = true
 allow_libvirt_bridge   = true
 
 [ai]
-endpoint        = "http://localhost:8080/v1"
-model           = "qwen2.5-coder:7b"
+# Point at Ollama (not LocalAI). Model uses Ollama name:tag format.
+endpoint        = "http://localhost:11434/v1"
+model           = "qwen3.5:2b"
 embed_model     = "nomic-embed-text"
 api_key         = ""
 enable_ollama   = true