diff --git a/raven/providers/common_models.py b/raven/providers/common_models.py
index d5da8c6..77e1333 100644
--- a/raven/providers/common_models.py
+++ b/raven/providers/common_models.py
@@ -82,6 +82,10 @@
         "deepseek/deepseek-v4-flash",
         "deepseek/deepseek-v4-pro",
     ],
+    "minimax": [
+        "minimax/MiniMax-M3",
+        "minimax/MiniMax-M2.7",
+    ],
     "minimax_global": [
         "minimax-global/MiniMax-M3",
         "minimax-global/MiniMax-M2.7",
diff --git a/raven/providers/rates.py b/raven/providers/rates.py
index e00365c..01de730 100644
--- a/raven/providers/rates.py
+++ b/raven/providers/rates.py
@@ -527,11 +527,29 @@ def _try_litellm_context_window(model: str, *, allow_import: bool = True) -> int
     return None
 
 
+def _registry_context_window(model: str) -> int | None:
+    """A current vendor window declared by the provider registry, or None."""
+    from raven.providers.registry import find_by_model, find_by_name, split_model_id
+
+    prefix, vendor_model = split_model_id(model)
+    if not prefix:
+        return None
+    spec = find_by_model(model)
+    if spec is None:
+        return None
+    source = find_by_name(spec.metadata_prefix) if spec.metadata_prefix else spec
+    if source is None:
+        return None
+    wanted = vendor_model.casefold()
+    return next((window for model_id, window in source.model_context_windows if model_id.casefold() == wanted), None)
+
+
 def resolve_context_window(model: str, *, allow_fetch: bool = True) -> int | None:
     """Return a model's real context window in tokens, or None.
 
-    LiteLLM's static metadata first, then OpenRouter's catalogue for ids that
-    name OpenRouter. The snapshot is deliberately not a source: a window sizes
+    Current model facts declared by the routing registry come first, then
+    LiteLLM's static metadata, then OpenRouter's catalogue for ids that name
+    OpenRouter. The snapshot is deliberately not a source: a window sizes
     trimming, so a community-maintained file that goes stale or wrong would
     shape the next request rather than cost a label. Unknown models return None
     so the caller keeps its configured default.
@@ -542,6 +560,10 @@ def resolve_context_window(model: str, *, allow_fetch: bool = True) -> int | Non
     behalf (see ``_try_litellm_context_window``) -- a caller cheap enough to
     pass this is cheap enough not to pay a fresh import either.
     """
+    window = _registry_context_window(model)
+    if window:
+        return window
+
     window = _try_litellm_context_window(model, allow_import=allow_fetch)
     if window:
         return window
diff --git a/raven/providers/registry.py b/raven/providers/registry.py
index 69d9c86..d6529c7 100644
--- a/raven/providers/registry.py
+++ b/raven/providers/registry.py
@@ -80,6 +80,10 @@ class ProviderSpec:
     # per-model param overrides, e.g. (("kimi-k2.5", {"temperature": 1.0}),)
     model_overrides: tuple[tuple[str, dict[str, Any]], ...] = ()
 
+    # Vendor model id to context window for current models the pinned LiteLLM
+    # metadata does not yet describe correctly.
+    model_context_windows: tuple[tuple[str, int], ...] = ()
+
     # OAuth-based providers (e.g., OpenAI Codex) don't use API keys
     is_oauth: bool = False  # if True, uses OAuth flow instead of API key
 
@@ -484,9 +488,11 @@ def claims(self, model: str) -> bool:
         detect_by_base_keyword="",
         strip_model_prefix=False,
         model_overrides=(),
+        model_context_windows=(("MiniMax-M3", 1_000_000), ("MiniMax-M2.7", 204_800)),
         # Needed by `provider test` and the wizard preflight, which probe
         # /v1/models before any LiteLLM call resolves an endpoint.
         default_api_base="https://api.minimax.io/v1",
+        default_model="minimax/MiniMax-M3",
     ),
     ProviderSpec(
         name="minimax_global",
diff --git a/tests/test_provider_catalog.py b/tests/test_provider_catalog.py
index c27e5e1..c9d2cee 100644
--- a/tests/test_provider_catalog.py
+++ b/tests/test_provider_catalog.py
@@ -119,6 +119,7 @@ def test_registry_and_schema_declare_the_same_providers() -> None:
     "zai",
     "dashscope",
     "groq",
+    "minimax",
     "minimax_global",
     "minimax_cn",
 ]
diff --git a/tests/test_provider_rates.py b/tests/test_provider_rates.py
index a88e558..65d061d 100644
--- a/tests/test_provider_rates.py
+++ b/tests/test_provider_rates.py
@@ -263,6 +263,19 @@ def test_a_litellm_mapped_window_comes_from_litellm_with_no_network(monkeypatch)
     assert counter["calls"] == 0
 
 
+@pytest.mark.parametrize(
+    ("model", "expected"),
+    [
+        ("minimax/MiniMax-M3", 1_000_000),
+        ("minimax/MiniMax-M2.7", 204_800),
+        ("minimax-global/MiniMax-M3", 1_000_000),
+        ("minimax-cn/MiniMax-M2.7", 204_800),
+    ],
+)
+def test_minimax_current_windows_come_from_the_provider_registry(model: str, expected: int) -> None:
+    assert resolve_context_window(model, allow_fetch=False) == expected
+
+
 def test_an_openrouter_window_falls_back_to_the_live_table(monkeypatch):
     _patch_litellm_info(monkeypatch, _litellm_miss)
     _patch_openrouter(monkeypatch, lambda req: _models_response(_DEEPSEEK_MODELS))