feat(runtime): select Gemma 4 by hardware profile

README.md

@@ -98,6 +98,16 @@ Exports are mapped to target formats without lock-in.
 ### Local-first execution
 Your hardware and selected profile determine latency/quality.
 
+Automatic selection uses the smallest effective local profile:
+
+- Fast: `gemma4:e2b-it-qat` for unknown or under-16-GB hardware and Intel Macs
+- Balanced: `gemma4:e4b-it-qat` for Apple Silicon with 16 GB or more
+- Advanced: explicit opt-in only; it is never selected automatically
+
+An explicit `CHROMEAI_OLLAMA_MODEL` override remains authoritative. If that
+model is unavailable locally, SelectPilot reports the missing model instead of
+silently substituting another model.
+
 ---
 
 ## Tiers

scripts/install-macos-local.sh

@@ -6,7 +6,7 @@ TEMPLATE="$ROOT/launchd/com.chromeai.nano.plist"
 DEST="${HOME}/Library/LaunchAgents/com.chromeai.nano.plist"
 HASH="$(shasum -a 256 "$ROOT/server/nano_server.py" | awk '{print $1}')"
 OLLAMA_BASE_URL="${CHROMEAI_OLLAMA_BASE_URL:-http://127.0.0.1:11434}"
-OLLAMA_MODEL="${CHROMEAI_OLLAMA_MODEL:-qwen2.5:0.5b}"
+OLLAMA_MODEL="${CHROMEAI_OLLAMA_MODEL:-gemma4:e2b-it-qat}"
 OLLAMA_EMBED_MODEL="${CHROMEAI_OLLAMA_EMBED_MODEL:-nomic-embed-text-v2-moe:latest}"
 RUN_DIR="${CHROMEAI_RUN_DIR:-${HOME}/Library/Application Support/SelectPilot/run}"
 LOG_DIR="${CHROMEAI_LOG_DIR:-${HOME}/Library/Logs/SelectPilot}"

server/ollama_client.py

@@ -65,29 +65,6 @@ def _build_markdown(summary: str, bullets: list[str], action_items: list[str] |
     return "\n".join(lines).strip() + "\n"
 
 
-GENERATION_MODEL_PREFERENCES = [
-    "llama3.2",
-    "llama3.1",
-    "qwen2.5",
-    "mistral",
-    "phi4",
-    "gemma3",
-    "glm-5-extended:latest",
-    "glm-5:cloud",
-    "gpt-oss:20b-cloud",
-    "qwen3.5:cloud",
-    "kimi-k2.5:cloud",
-    "minimax-m2.5:cloud",
-    "deepseek-v3.2:cloud",
-]
-
-EMBED_MODEL_PREFERENCES = [
-    "nomic-embed-text-v2-moe:latest",
-    "nomic-embed-text",
-    "mxbai-embed-large",
-]
-
-
 @dataclass(frozen=True)
 class OllamaConfig:
     base_url: str
@@ -103,7 +80,7 @@ class OllamaError(RuntimeError):
 class OllamaClient:
     def __init__(self, config: OllamaConfig | None = None):
         if config is None:
-            default_generation_model = "llama3.2"
+            default_generation_model = "gemma4:e2b-it-qat"
             default_embed_model = "nomic-embed-text-v2-moe:latest"
             runtime_profile = os.environ.get("CHROMEAI_RUNTIME_PROFILE", "auto")
 
@@ -209,17 +186,6 @@ def _model_names(self, local_only: bool = False) -> list[str]:
                 models.append(str(name))
         return models
 
-    def _resolve_model(self, requested: str, preferences: list[str], models: list[str]) -> str:
-        if requested in models:
-            return requested
-        for candidate in preferences:
-            for model in models:
-                if model == candidate or model.startswith(f"{candidate}:"):
-                    return model
-        if models:
-            return models[0]
-        return requested
-
     def _request_json(self, path: str, payload: dict[str, Any] | None = None) -> Any:
         url = urljoin(self.config.base_url + "/", path.lstrip("/"))
         data = None if payload is None else json.dumps(payload).encode("utf-8")
@@ -243,12 +209,12 @@ def tags(self) -> dict[str, Any]:
         return self._request_json("/api/tags", None)
 
     def active_generation_model(self, models: list[str] | None = None) -> str:
-        models = models if models is not None else self._model_names()
-        return self._resolve_model(self.config.model, GENERATION_MODEL_PREFERENCES, models)
+        # The configured hardware profile is authoritative. A different local
+        # model must never be selected implicitly when the requested one is absent.
+        return self.config.model
 
     def active_embedding_model(self, models: list[str] | None = None) -> str:
-        models = models if models is not None else self._model_names()
-        return self._resolve_model(self.config.embed_model, EMBED_MODEL_PREFERENCES, models)
+        return self.config.embed_model
 
     def health(self) -> dict[str, Any]:
         try:

server/runtime_profiles.py

@@ -24,8 +24,8 @@ class RuntimeProfile:
     "fast": RuntimeProfile(
         key="fast",
         label="Fast",
-        description="Smallest viable local profile for structured extraction and low-latency summaries.",
-        generation_model="qwen2.5:0.5b",
+        description="Memory-safe local profile for structured extraction and low-latency summaries.",
+        generation_model="gemma4:e2b-it-qat",
         embedding_model="nomic-embed-text-v2-moe:latest",
         target_latency="1-4s",
         intended_for="Selected-text extraction, action briefs, and quick summaries.",
@@ -34,8 +34,8 @@ class RuntimeProfile:
     "balanced": RuntimeProfile(
         key="balanced",
         label="Balanced",
-        description="Higher quality local profile for rewrite and general-purpose browser transforms.",
-        generation_model="qwen2.5:3b",
+        description="Efficient local profile for higher-quality browser transforms on 16 GB or more.",
+        generation_model="gemma4:e4b-it-qat",
         embedding_model="nomic-embed-text-v2-moe:latest",
         target_latency="2-6s",
         intended_for="Daily use when you want better quality without drifting into heavy models.",
@@ -132,4 +132,3 @@ def build_bootstrap_commands(profile_key: str, project_root: str | Path) -> dict
         "generation_model": profile.generation_model,
         "embedding_model": profile.embedding_model,
     }
-

tests/server/test_runtime_profiles.py

@@ -2,6 +2,7 @@
 
 import sys
 import unittest
+from unittest.mock import patch
 from pathlib import Path
 
 ROOT = Path(__file__).resolve().parents[2]
@@ -10,13 +11,36 @@
     sys.path.insert(0, str(SERVER_DIR))
 
 from runtime_profiles import build_bootstrap_commands, get_runtime_profile, recommend_runtime_profile  # noqa: E402
+from ollama_client import OllamaClient  # noqa: E402
 
 
 class RuntimeProfileTests(unittest.TestCase):
     def test_unknown_profile_falls_back_to_fast(self) -> None:
         profile = get_runtime_profile("missing")
         self.assertEqual(profile.key, "fast")
-        self.assertEqual(profile.generation_model, "qwen2.5:0.5b")
+        self.assertEqual(profile.generation_model, "gemma4:e2b-it-qat")
+
+    def test_low_memory_hardware_uses_memory_safe_gemma_profile(self) -> None:
+        recommendation = recommend_runtime_profile(
+            {"machine": "arm64", "memory_gb": 8, "platform": "darwin", "cpu_count": 8}
+        )
+        profile = get_runtime_profile(recommendation["recommended_profile"])
+        self.assertEqual(profile.key, "fast")
+        self.assertEqual(profile.generation_model, "gemma4:e2b-it-qat")
+
+    def test_16_gb_apple_silicon_uses_balanced_gemma_profile(self) -> None:
+        recommendation = recommend_runtime_profile(
+            {"machine": "arm64", "memory_gb": 16, "platform": "darwin", "cpu_count": 10}
+        )
+        profile = get_runtime_profile(recommendation["recommended_profile"])
+        self.assertEqual(profile.key, "balanced")
+        self.assertEqual(profile.generation_model, "gemma4:e4b-it-qat")
+
+    def test_intel_hardware_remains_on_memory_safe_profile(self) -> None:
+        recommendation = recommend_runtime_profile(
+            {"machine": "x86_64", "memory_gb": 64, "platform": "darwin", "cpu_count": 12}
+        )
+        self.assertEqual(recommendation["recommended_profile"], "fast")
 
     def test_auto_recommendation_prefers_balanced_on_large_machines(self) -> None:
         recommendation = recommend_runtime_profile(
@@ -28,6 +52,26 @@ def test_bootstrap_command_contains_profile(self) -> None:
         command = build_bootstrap_commands("balanced", ROOT)
         self.assertIn("--profile balanced", command["command"])
 
+    def test_ollama_client_uses_explicit_profile_without_model_fallback(self) -> None:
+        with patch.dict("os.environ", {"CHROMEAI_RUNTIME_PROFILE": "balanced"}, clear=True):
+            client = OllamaClient()
+
+        self.assertEqual(client.config.model, "gemma4:e4b-it-qat")
+        self.assertEqual(client.active_generation_model(["qwen2.5:0.5b"]), "gemma4:e4b-it-qat")
+
+    def test_explicit_model_override_remains_authoritative(self) -> None:
+        with patch.dict(
+            "os.environ",
+            {
+                "CHROMEAI_RUNTIME_PROFILE": "balanced",
+                "CHROMEAI_OLLAMA_MODEL": "custom-local:model",
+            },
+            clear=True,
+        ):
+            client = OllamaClient()
+
+        self.assertEqual(client.config.model, "custom-local:model")
+
 
 if __name__ == "__main__":
     unittest.main()