feat: default new MiniMax text agents to M3.1 Flash

.env.example

@@ -30,7 +30,13 @@ XAI_API_KEY=
 GOOGLE_API_KEY=
 GOOGLE_GENAI_USE_VERTEXAI=false
 #GOOGLE_CLOUD_PROJECT=
+# Direct API: new text requests use MiniMax-M3.1-Flash; saved M3 requests stay M3.
+MINIMAX_API_KEY=
+MINIMAX_BASE_URL=https://api.minimax.io/anthropic
+# Existing deployments may keep these fallback settings. Account/model access
+# depends on the configured endpoint and plan; no domestic plan entitlement is implied.
 MINIMAX_PLAN_API_KEY=
+#MINIMAX_PLAN_BASE_URL=
 MIMO_PLAN_API_KEY=
 
 # OpenAI Compatible Provider

intentkit/config/config.py

@@ -155,7 +155,13 @@ def __init__(self) -> None:
         self.google_cloud_project: str | None = self.load("GOOGLE_CLOUD_PROJECT")
         self.deepseek_api_key: str | None = self.load("DEEPSEEK_API_KEY")
         self.xai_api_key: str | None = self.load("XAI_API_KEY")
-        self.minimax_plan_api_key: str | None = self.load("MINIMAX_PLAN_API_KEY")
+        self.minimax_base_url: str = self.load(
+            "MINIMAX_BASE_URL",
+            self.load("MINIMAX_PLAN_BASE_URL", "https://api.minimax.io/anthropic"),
+        )
+        self.minimax_plan_api_key: str | None = self.load(
+            "MINIMAX_API_KEY"
+        ) or self.load("MINIMAX_PLAN_API_KEY")
         self.mimo_plan_api_key: str | None = self.load("MIMO_PLAN_API_KEY")
         self.openrouter_api_key: str | None = self.load("OPENROUTER_API_KEY")
         # OpenAI Compatible provider

intentkit/models/llm.py

@@ -242,6 +242,10 @@ class LLMModelInfo(BaseModel):
         ),
     )
     enabled: bool = Field(default=True)
+    selectable: bool = Field(
+        default=True,
+        description="Offer in model pickers; false retains explicit saved model requests.",
+    )
     input_price: Decimal  # Price per 1M input tokens in USD
     cached_input_price: Decimal | None = None  # Price per 1M cached input tokens in USD
     output_price: Decimal  # Price per 1M output tokens in USD
@@ -424,7 +428,7 @@ async def get_all(cls) -> list["LLMModelInfo"]:
 
         Kept ``async`` (despite no I/O) so the awaited call sites stay stable.
         """
-        return list(AVAILABLE_MODELS.values())
+        return [model for model in AVAILABLE_MODELS.values() if model.selectable]
 
     async def calculate_cost(
         self, input_tokens: int, output_tokens: int, cached_input_tokens: int = 0
@@ -846,9 +850,20 @@ async def create_instance(self) -> BaseChatModel:
             "app_categories": ["cloud-agent"],
         }
 
-        # Agent efforts use OpenRouter's own enum, so the clamped value goes on
-        # the wire as-is. "none" is deliberately forwarded: OpenRouter uses it
-        # to disable reasoning on hybrid upstream models.
+        # Flash always thinks; reject unsupported explicit values rather than
+        # silently turning a request for disabled thinking into a thinking run.
+        if (
+            info.api_id == "minimax/minimax-m3.1-flash:free"
+            and self.reasoning_effort
+            in {
+                "none",
+                "minimal",
+            }
+        ):
+            raise ValueError(
+                "MiniMax-M3.1-Flash requires low/medium/high/xhigh/max effort"
+            )
+        # Other models retain OpenRouter's normal effort clamping.
         effort = info.resolve_reasoning_effort(self.reasoning_effort)
         if effort is not None:
             kwargs["reasoning"] = {"effort": effort}
@@ -929,20 +944,26 @@ async def create_instance(self) -> BaseChatModel:
         kwargs: dict[str, Any] = {
             "model": info.api_id,
             "api_key": config.minimax_plan_api_key,
-            "base_url": "https://api.minimax.io/anthropic",
+            "base_url": config.minimax_base_url,
             "timeout": info.timeout,
             "max_retries": 3,
         }
 
-        # MiniMax's Anthropic-compatible endpoint defaults to thinking
-        # DISABLED (unlike its OpenAI-compatible one), so always send the
-        # toggle. It only understands "adaptive" and "disabled". Note the
-        # generic AnthropicCompatibleLLM below deliberately sends no thinking
-        # config: arbitrary endpoints disagree on the accepted shape.
-        effort = info.resolve_reasoning_effort(self.reasoning_effort)
-        kwargs["thinking"] = {
-            "type": "adaptive" if effort and effort != "none" else "disabled"
-        }
+        if info.api_id == "MiniMax-M3.1-Flash":
+            effort = self.reasoning_effort or info.reasoning_effort or "max"
+            if effort not in {"low", "medium", "high", "xhigh", "max"}:
+                raise ValueError(
+                    "MiniMax-M3.1-Flash requires low/medium/high/xhigh/max effort"
+                )
+            kwargs["thinking"] = {"type": "adaptive"}
+            kwargs["output_config"] = {"effort": effort}
+        elif info.api_id != "MiniMax-M2.7":
+            # M2.7 keeps its always-on default. M3 ignores effort depth;
+            # keep the saved M3 agent's thinking toggle.
+            effort = info.resolve_reasoning_effort(self.reasoning_effort)
+            kwargs["thinking"] = {
+                "type": "adaptive" if effort and effort != "none" else "disabled"
+            }
 
         return ChatAnthropic(**kwargs)
 

intentkit/models/llm.yaml

@@ -133,12 +133,33 @@
   reasoning_levels: ["none", "high"]
   timeout: 300
 
+# Gateway metadata follows OpenRouter's Minimax endpoint, not direct API prices.
+# https://openrouter.ai/api/v1/models/minimax/minimax-m3.1-flash:free/endpoints
+- id: "minimax/minimax-flash:free"
+  api_id: "minimax/minimax-m3.1-flash:free"
+  name: "MiniMax M3.1 Flash (free)"
+  provider: openrouter
+  origin_provider: "minimax"
+  enabled: true
+  input_price: "0"
+  output_price: "0"
+  price_level: 1
+  context_length: 1000000
+  output_length: 524288
+  intelligence: 5
+  speed: 4
+  supports_image_input: true
+  supports_audio_input: false
+  supports_video_input: true
+  reasoning_effort: "max"
+  reasoning_levels: ["low", "medium", "high", "xhigh", "max"]
+  timeout: 300
+
 - id: "minimax/minimax"
   api_id: "minimax/minimax-m3"
   name: "MiniMax M3"
   provider: openrouter
   origin_provider: "minimax/fp8"   # lock OpenRouter upstream
-  legacy_ids: ["minimax/minimax-m2-her"]
   enabled: true
   input_price: "0.3"
   cached_input_price: "0.06"
@@ -848,10 +869,68 @@
   reasoning_levels: ["low", "medium", "high", "xhigh"]
   timeout: 180
 
+# https://platform.minimax.io/docs/api-reference/text-anthropic-api
+# https://platform.minimax.io/docs/pricing/overview
+- id: "minimax-flash"
+  api_id: "MiniMax-M3.1-Flash"
+  name: "MiniMax M3.1 Flash"
+  provider: minimax
+  enabled: true
+  input_price: "0.15"
+  cached_input_price: "0.015"
+  output_price: "0.60"
+  price_level: 1
+  context_length: 1000000
+  output_length: 524288
+  intelligence: 5
+  speed: 4
+  supports_image_input: true
+  supports_audio_input: false
+  supports_video_input: true
+  reasoning_effort: "max"
+  reasoning_levels: ["low", "medium", "high", "xhigh", "max"]
+  timeout: 300
+
+# Explicit legacy requests remain routable but are not ordinary picker choices.
+- id: "MiniMax-M2.7"
+  api_id: "MiniMax-M2.7"
+  name: "MiniMax M2.7 (legacy)"
+  provider: minimax
+  selectable: false
+  input_price: "0.3"
+  cached_input_price: "0.06"
+  output_price: "1.2"
+  context_length: 204800
+  output_length: 204800
+  intelligence: 4
+  speed: 4
+  supports_image_input: false
+  supports_audio_input: false
+  supports_video_input: false
+  timeout: 300
+
+- id: "minimax/minimax-m2-her"
+  api_id: "minimax/minimax-m2-her"
+  name: "MiniMax M2-her (legacy)"
+  provider: openrouter
+  selectable: false
+  input_price: "0.3"
+  cached_input_price: "0.03"
+  output_price: "1.2"
+  context_length: 65536
+  output_length: 2048
+  intelligence: 4
+  speed: 4
+  supports_image_input: false
+  supports_audio_input: false
+  supports_video_input: false
+  timeout: 300
+
 - id: "minimax"
   api_id: "MiniMax-M3"
   name: "MiniMax M3"
   provider: minimax
+  legacy_ids: ["minimax:MiniMax-M3"]
   enabled: true
   input_price: "0.3"
   cached_input_price: "0.06"

intentkit/models/llm_picker.py

@@ -68,7 +68,7 @@ def pick_summarize_model() -> str:
         ("gpt-luna", LLMProvider.OPENAI),
         ("grok", LLMProvider.XAI),
         ("deepseek-flash", LLMProvider.DEEPSEEK),
-        ("minimax", LLMProvider.MINIMAX),
+        ("minimax-flash", LLMProvider.MINIMAX),
         ("mimo", LLMProvider.MIMO_PLAN),
     ]
     return _first_configured(order, lite_compatible=True)
@@ -82,8 +82,8 @@ def pick_default_model() -> str:
     """
     order: list[tuple[str, LLMProvider]] = [
         ("gemini-flash", LLMProvider.GOOGLE),
-        ("minimax", LLMProvider.MINIMAX),
-        ("minimax/minimax", LLMProvider.OPENROUTER),
+        ("minimax-flash", LLMProvider.MINIMAX),
+        ("minimax/minimax-flash:free", LLMProvider.OPENROUTER),
         ("gpt-luna", LLMProvider.OPENAI),
         ("grok", LLMProvider.XAI),
         ("deepseek-flash", LLMProvider.DEEPSEEK),
@@ -105,7 +105,7 @@ def pick_lead_model() -> str:
         ("gemini-flash", LLMProvider.GOOGLE),
         ("gpt-luna", LLMProvider.OPENAI),
         ("grok", LLMProvider.XAI),
-        ("minimax", LLMProvider.MINIMAX),
+        ("minimax-flash", LLMProvider.MINIMAX),
         ("mimo", LLMProvider.MIMO_PLAN),
     ]
     return _first_configured(order, fallback=_DEFAULT_FALLBACK_MODEL)
@@ -120,7 +120,7 @@ def pick_lite_model() -> str:
         # Luna is OpenAI's cheapest tier; glm/deepseek above are still cheaper.
         ("gpt-luna", LLMProvider.OPENAI),
         ("grok", LLMProvider.XAI),
-        ("minimax", LLMProvider.MINIMAX),
+        ("minimax-flash", LLMProvider.MINIMAX),
         ("mimo", LLMProvider.MIMO_PLAN),
     ]
     return _first_configured(
@@ -136,7 +136,7 @@ def pick_smartest_model() -> str:
         ("gpt-sol", LLMProvider.OPENAI),
         ("grok", LLMProvider.XAI),
         ("deepseek-pro", LLMProvider.DEEPSEEK),
-        ("minimax", LLMProvider.MINIMAX),
+        ("minimax-flash", LLMProvider.MINIMAX),
         ("mimo-pro", LLMProvider.MIMO_PLAN),
     ]
     return _first_configured(order, fallback=_DEFAULT_FALLBACK_MODEL)
@@ -150,7 +150,7 @@ def pick_fastest_model() -> str:
         ("gpt-luna", LLMProvider.OPENAI),
         ("grok", LLMProvider.XAI),
         ("deepseek-flash", LLMProvider.DEEPSEEK),
-        ("minimax", LLMProvider.MINIMAX),
+        ("minimax-flash", LLMProvider.MINIMAX),
         ("mimo", LLMProvider.MIMO_PLAN),
     ]
     return _first_configured(
@@ -178,7 +178,7 @@ def pick_writing_model() -> str:
         ("anthropic/claude-sonnet", LLMProvider.OPENROUTER),
         ("gemini-pro", LLMProvider.GOOGLE),
         ("gpt-sol", LLMProvider.OPENAI),
-        ("minimax", LLMProvider.MINIMAX),
+        ("minimax-flash", LLMProvider.MINIMAX),
         ("deepseek-pro", LLMProvider.DEEPSEEK),
         ("grok", LLMProvider.XAI),
         ("mimo-pro", LLMProvider.MIMO_PLAN),
@@ -190,7 +190,7 @@ def pick_chinese_writing_model() -> str:
     """Pick the best model for Chinese writing (Chinese-native models first)."""
     order: list[tuple[str, LLMProvider]] = [
         ("qwen/qwen-max", LLMProvider.OPENROUTER),
-        ("minimax", LLMProvider.MINIMAX),
+        ("minimax-flash", LLMProvider.MINIMAX),
         ("mimo-pro", LLMProvider.MIMO_PLAN),
         ("deepseek-pro", LLMProvider.DEEPSEEK),
         ("gemini-pro", LLMProvider.GOOGLE),
@@ -208,7 +208,7 @@ def pick_finance_model() -> str:
         ("gemini-pro", LLMProvider.GOOGLE),
         ("gpt-sol", LLMProvider.OPENAI),
         ("grok", LLMProvider.XAI),
-        ("minimax", LLMProvider.MINIMAX),
+        ("minimax-flash", LLMProvider.MINIMAX),
         ("mimo-pro", LLMProvider.MIMO_PLAN),
     ]
     return _first_configured(order, fallback=_DEFAULT_FALLBACK_MODEL)
@@ -222,7 +222,7 @@ def pick_search_model() -> str:
         ("gpt-terra", LLMProvider.OPENAI),
         ("x-ai/grok", LLMProvider.OPENROUTER),
         ("deepseek-flash", LLMProvider.DEEPSEEK),
-        ("minimax", LLMProvider.MINIMAX),
+        ("minimax-flash", LLMProvider.MINIMAX),
         ("mimo", LLMProvider.MIMO_PLAN),
     ]
     return _first_configured(order, fallback=_DEFAULT_FALLBACK_MODEL)
@@ -236,7 +236,7 @@ def pick_broadest_knowledge_model() -> str:
         ("gpt-sol", LLMProvider.OPENAI),
         ("grok", LLMProvider.XAI),
         ("deepseek-pro", LLMProvider.DEEPSEEK),
-        ("minimax", LLMProvider.MINIMAX),
+        ("minimax-flash", LLMProvider.MINIMAX),
         ("mimo-pro", LLMProvider.MIMO_PLAN),
     ]
     return _first_configured(order, fallback=_DEFAULT_FALLBACK_MODEL)
@@ -253,16 +253,24 @@ def pick_long_context_model() -> str:
         ("deepseek/deepseek-flash", LLMProvider.OPENROUTER),
         ("deepseek-flash", LLMProvider.DEEPSEEK),
         ("gpt-luna", LLMProvider.OPENAI),
-        ("minimax", LLMProvider.MINIMAX),
+        ("minimax-flash", LLMProvider.MINIMAX),
         ("mimo", LLMProvider.MIMO_PLAN),
     ]
     return _first_configured(order)
 
 
 def list_available_model_ids() -> list[str]:
-    """Return the sorted, distinct model IDs available in this deployment.
+    """Return distinct model IDs, alphabetically except MiniMax Flash before M3.
 
     Reflects the providers configured at process start (``AVAILABLE_MODELS``),
     which is fixed for the lifetime of a deployment.
     """
-    return sorted({model.id for model in AVAILABLE_MODELS.values()})
+    ids = sorted({model.id for model in AVAILABLE_MODELS.values() if model.selectable})
+    for flash, m3 in (
+        ("minimax-flash", "minimax"),
+        ("minimax/minimax-flash:free", "minimax/minimax"),
+    ):
+        if flash in ids and m3 in ids:
+            first, second = sorted((ids.index(flash), ids.index(m3)))
+            ids[first], ids[second] = flash, m3
+    return ids

tests/core/test_llm.py

@@ -351,7 +351,8 @@ def test_catalog_ids_are_versionless_series_ids():
 
     rows = _catalog_rows()
     for row in rows:
-        assert not re.search(r"\d", row["id"]), f"{row['id']} carries a version"
+        if row.get("selectable", True):
+            assert not re.search(r"\d", row["id"]), f"{row['id']} carries a version"
         api_id = row.get("api_id", row["id"])
         assert api_id, row["id"]
         if api_id.startswith("~"):

tests/core/test_origin_provider.py

@@ -104,7 +104,9 @@ def test_all_openrouter_catalog_entries_pin_origin_provider():
     unpinned = [
         row["id"]
         for row in rows
-        if row["provider"] == "openrouter" and not row.get("origin_provider")
+        if row["provider"] == "openrouter"
+        and row.get("selectable", True)
+        and not row.get("origin_provider")
     ]
     assert unpinned == ["openrouter/free"]
 
@@ -118,6 +120,7 @@ async def fake_get(model_id: str) -> LLMModelInfo:
         return info
 
     monkeypatch.setattr(LLMModelInfo, "get", staticmethod(fake_get))
+    monkeypatch.setattr("intentkit.models.llm.config.openrouter_api_key", "test-key")
 
     llm = OpenRouterLLM(model_name=info.id, info=info)
     instance = await llm.create_instance()
@@ -138,6 +141,7 @@ async def fake_get(model_id: str) -> LLMModelInfo:
         return info
 
     monkeypatch.setattr(LLMModelInfo, "get", staticmethod(fake_get))
+    monkeypatch.setattr("intentkit.models.llm.config.openrouter_api_key", "test-key")
 
     llm = OpenRouterLLM(model_name=info.id, info=info)
     instance = await llm.create_instance()

tests/core/test_reasoning_effort.py

@@ -237,6 +237,7 @@ async def test_minimax_always_sends_thinking_toggle():
         patch.object(LLMModelInfo, "get", return_value=info),
     ):
         mock_config.minimax_plan_api_key = "test-key"
+        mock_config.minimax_base_url = "https://api.minimax.io/anthropic"
         on = await MiniMaxLLM(model_name=info.id, info=info).create_instance()
         off = await MiniMaxLLM(
             model_name=info.id, reasoning_effort="none", info=info