$(cat <<EOF

Fix model ID format for Cerebras and Silicon Flow providers.

Cerebras API expects bare model IDs (e.g. "qwen-3-235b-a22b-instruct-2507")
not the full "provider/model" format. Strip cerebras/ prefix before sending.

Silicon Flow API similarly expects bare model IDs. Strip silicon/ prefix.

Also updated REQUESTED_PROVIDER_MODELS and MODEL_CAPABILITIES to use
correct ref format (cerebras/qwen-3-235b-a22b-instruct-2507, etc.).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
EOF
)

api/routes.py

@@ -41,10 +41,10 @@ REQUESTED_PROVIDER_MODELS = [
     "nvidia_nim/z-ai/glm4.7",
     "nvidia_nim/bytedance/seed-oss-36b-instruct",
     "nvidia_nim/mistralai/mistral-nemotron",
-    # Cerebras models
-    "cerebras/qwen/qwen-3-235b-a22b-instruct-2507",
-    "cerebras/z-ai/glm4.7",
-    # Silicon Flow models
+    # Cerebras models (uses bare model IDs on the API)
+    "cerebras/qwen-3-235b-a22b-instruct-2507",
+    "cerebras/zai-glm-4.7",
+    # Silicon Flow models (uses bare model IDs on the API)
     "silicon/Qwen/Qwen3.6-35B-A3B",
     "silicon/Qwen/Qwen3.6-27B",
     "silicon/Qwen/Qwen3.5-35B-A3B",

core/model_capabilities.py

@@ -132,10 +132,10 @@ MODEL_CAPABILITIES: dict[str, ModelCapabilities] = {
         priority=60,
     ),
     # Cerebras models
-    "cerebras/qwen/qwen-3-235b-a22b-instruct-2507": ModelCapabilities(
+    "cerebras/qwen-3-235b-a22b-instruct-2507": ModelCapabilities(
         provider_id="cerebras",
-        model_id="qwen/qwen-3-235b-a22b-instruct-2507",
-        model_ref="cerebras/qwen/qwen-3-235b-a22b-instruct-2507",
+        model_id="qwen-3-235b-a22b-instruct-2507",
+        model_ref="cerebras/qwen-3-235b-a22b-instruct-2507",
         coding=True,
         reasoning=True,
         general_text=True,
@@ -143,10 +143,10 @@ MODEL_CAPABILITIES: dict[str, ModelCapabilities] = {
         speed="slow",
         priority=85,
     ),
-    "cerebras/z-ai/glm4.7": ModelCapabilities(
+    "cerebras/zai-glm-4.7": ModelCapabilities(
         provider_id="cerebras",
-        model_id="z-ai/glm4.7",
-        model_ref="cerebras/z-ai/glm4.7",
+        model_id="zai-glm-4.7",
+        model_ref="cerebras/zai-glm-4.7",
         coding=True,
         reasoning=True,
         general_text=True,

providers/cerebras/client.py

@@ -12,6 +12,16 @@ from providers.openai_compat import OpenAIChatTransport
 class CerebrasProvider(OpenAIChatTransport):
     """Cerebras provider using OpenAI-compatible /chat/completions."""
 
+    # Mapping of proxy model refs to Cerebras API model IDs.
+    # The proxy uses full refs like "cerebras/qwen-3-235b-a22b-instruct-2507"
+    # but Cerebras API expects bare model IDs like "qwen-3-235b-a22b-instruct-2507".
+    CEREBRAS_MODEL_MAP: dict[str, str] = {
+        "qwen-3-235b-a22b-instruct-2507": "qwen-3-235b-a22b-instruct-2507",
+        "zai-glm-4.7": "zai-glm-4.7",
+        "cerebras/qwen-3-235b-a22b-instruct-2507": "qwen-3-235b-a22b-instruct-2507",
+        "cerebras/z-ai/glm4.7": "zai-glm-4.7",
+    }
+
     def __init__(self, config: ProviderConfig, *, settings: Settings):
         base_url = (config.base_url or CEREBRAS_DEFAULT_BASE).rstrip("/")
         if not base_url.endswith("/v1"):
@@ -33,4 +43,11 @@ class CerebrasProvider(OpenAIChatTransport):
             if thinking
             else ReasoningReplayMode.DISABLED
         )
-        return build_base_request_body(request, reasoning_replay=reasoning_replay)
+        body = build_base_request_body(request, reasoning_replay=reasoning_replay)
+        # Strip cerebras/ prefix so the API gets the bare model ID
+        model = body.get("model", "")
+        if model in self.CEREBRAS_MODEL_MAP:
+            body["model"] = self.CEREBRAS_MODEL_MAP[model]
+        elif model.startswith("cerebras/"):
+            body["model"] = model[len("cerebras/") :]
+        return body

providers/silicon/client.py

@@ -33,4 +33,9 @@ class SiliconProvider(OpenAIChatTransport):
             if thinking
             else ReasoningReplayMode.DISABLED
         )
-        return build_base_request_body(request, reasoning_replay=reasoning_replay)
+        body = build_base_request_body(request, reasoning_replay=reasoning_replay)
+        # Strip silicon/ prefix so the API gets the bare model ID
+        model = body.get("model", "")
+        if model.startswith("silicon/"):
+            body["model"] = model[len("silicon/") :]
+        return body