diff --git a/src/fabledassistant/routes/settings.py b/src/fabledassistant/routes/settings.py
index 33d25cd..2c6a8a6 100644
--- a/src/fabledassistant/routes/settings.py
+++ b/src/fabledassistant/routes/settings.py
@@ -64,6 +64,15 @@ async def update_settings_route():
     if not isinstance(data, dict):
         return jsonify({"error": "Expected a JSON object"}), 400
 
+    # Normalize model names to lowercase before validation. Ollama's /api/tags
+    # preserves whatever casing was used at pull time, but /api/chat rejects
+    # mixed-case tags — lowercasing here keeps the two paths consistent and
+    # means the stored setting is always in a form Ollama will actually accept.
+    _MODEL_KEYS = frozenset({"default_model", "background_model"})
+    for _key in _MODEL_KEYS:
+        if _key in data and data[_key]:
+            data[_key] = str(data[_key]).lower()
+
     if "default_model" in data:
         installed = await get_installed_models()
         if data["default_model"]:
@@ -74,7 +83,6 @@ async def update_settings_route():
     # Empty string for model keys means "reset to system default".
     # Delete the DB row so get_setting() falls back to Config defaults
     # rather than returning "" and breaking model resolution everywhere.
-    _MODEL_KEYS = frozenset({"default_model", "background_model"})
     to_save = {}
     for k, v in data.items():
         str_v = str(v)
diff --git a/src/fabledassistant/services/llm.py b/src/fabledassistant/services/llm.py
index acd79da..288df68 100644
--- a/src/fabledassistant/services/llm.py
+++ b/src/fabledassistant/services/llm.py
@@ -78,7 +78,14 @@ RAG_AUTO_SNIPPET = 4000
 
 
 async def get_installed_models() -> set[str]:
-    """Return set of installed Ollama model names (with and without :latest)."""
+    """Return set of installed Ollama model names (with and without :latest).
+
+    Tags are normalized to lowercase. Ollama stores whatever casing was used at
+    pull time (``ollama pull gemma3:12B`` keeps the ``B``), but inference calls
+    against the capitalized tag can 400 — the two code paths are inconsistent.
+    Lowercasing on read avoids exposing that mixed-case name in the UI and
+    keeps the validation check below from accepting a form Ollama will reject.
+    """
     try:
         async with httpx.AsyncClient(timeout=10.0) as client:
             resp = await client.get(f"{Config.OLLAMA_URL}/api/tags")
@@ -86,7 +93,7 @@ async def get_installed_models() -> set[str]:
             data = resp.json()
             names: set[str] = set()
             for m in data.get("models", []):
-                name = m["name"]
+                name = m["name"].lower()
                 names.add(name)
                 if name.endswith(":latest"):
                     names.add(name.removesuffix(":latest"))
@@ -98,6 +105,9 @@ async def get_installed_models() -> set[str]:
 
 async def ensure_model(model: str) -> None:
     """Check if model exists in Ollama, pull if missing."""
+    # Match the lowercase normalization in get_installed_models so a legacy
+    # mixed-case setting doesn't force a spurious re-pull at startup.
+    model = model.lower()
     try:
         installed = await get_installed_models()
         if model in installed or f"{model}:latest" in installed: