fix: add initial support for magistral models

ErikBjare · ErikBjare · commit 59dce119f425 · 2025-09-05T21:03:40.000+02:00
diff --git a/gptme/llm/llm_openai.py b/gptme/llm/llm_openai.py
@@ -604,11 +604,17 @@ def _prepare_messages_for_api(
         messages = list(_prep_o1(messages))
 
     # without this, deepseek-chat and reasoner can start outputting gibberish after tool calls
-    # similarly, kimi-k2-instruct doesn't acknowledge tool responses/system messages without it
+    # similarly, kimi-k2-instruct doesn't acknowledge tool responses/system messages without it, same with magistral
     # it probably applies to more models/providers, we should figure out which and perhaps make it default behavior
+    # TODO: it seems to apply to a lot of reasoning models, should maybe be default behavior for all reasoning models?
     if any(
         m in model_meta.model
-        for m in ["deepseek-reasoner", "deepseek-chat", "kimi-k2-instruct"]
+        for m in [
+            "deepseek-reasoner",
+            "deepseek-chat",
+            "kimi-k2-instruct",
+            "magistral-medium-2506",
+        ]
     ):
         messages = list(_prep_deepseek_reasoner(messages))
 
diff --git a/gptme/llm/models.py b/gptme/llm/models.py
@@ -265,6 +265,14 @@ class _ModelDictMeta(TypedDict):
         },
     },
     "openrouter": {
+        "mistralai/magistral-medium-2506": {
+            "context": 41_000,
+            "max_output": 40_000,
+            "price_input": 2,
+            "price_output": 5,
+            # "supports_vision": True,
+            "supports_reasoning": True,
+        },
         "anthropic/claude-3.5-sonnet": {
             "context": 200_000,
             "max_output": 8192,
@@ -388,13 +396,13 @@ def _get_models_for_provider(
     provider: Provider, dynamic_fetch: bool = True
 ) -> list[ModelMeta]:
     """Get models for a specific provider, with optional dynamic fetching."""
+    from . import get_available_models  # fmt: skip
+
     models_to_show = []
 
     # Try dynamic fetching first for supported providers
     if dynamic_fetch and provider == "openrouter":
         try:
-            from . import get_available_models
-
             dynamic_models = get_available_models(provider)
             models_to_show = dynamic_models
         except Exception:
