diff --git a/src/llm_core.py b/src/llm_core.py
index 7af9c2725..e8331279f 100644
--- a/src/llm_core.py
+++ b/src/llm_core.py
@@ -677,6 +677,8 @@ def _detect_provider(url: str) -> str:
     from src.copilot import is_copilot_base
     if is_copilot_base(url):
         return "copilot"
+    if _host_match(url, "cerebras.ai"):
+        return "cerebras"
     if _host_match(url, "mistral.ai"):
         return "mistral"
     return "openai"
@@ -763,6 +765,8 @@ def _provider_label(url: str) -> str:
     if is_chatgpt_subscription_base(url): return "ChatGPT Subscription"
     from src.copilot import is_copilot_base
     if is_copilot_base(url): return "GitHub Copilot"
+    if _host_match(url, "cerebras.ai"):
+        return "cerebras"
     if _host_match(url, "mistral.ai"): return "Mistral"
     if _host_match(url, "deepseek.com"): return "DeepSeek"
     if _host_match(url, "nvidia.com"): return "NVIDIA"
diff --git a/tests/test_cerebras_cache_affinity.py b/tests/test_cerebras_cache_affinity.py
new file mode 100644
index 000000000..a5144feda
--- /dev/null
+++ b/tests/test_cerebras_cache_affinity.py
@@ -0,0 +1,27 @@
+"""Regression test for issue #4640.
+
+Cerebras endpoints must not receive llama.cpp-specific fields
+(session_id, cache_prompt) even when endpoint_kind is misconfigured as 'local'.
+"""
+import importlib
+
+
+def test_detect_provider_recognizes_cerebras():
+    """_detect_provider should return 'cerebras' for api.cerebras.ai URLs."""
+    llm_core = importlib.import_module("src.llm_core")
+    assert llm_core._detect_provider("https://api.cerebras.ai/v1") == "cerebras"
+
+
+def test_cerebras_not_self_hosted():
+    """_is_self_hosted_openai_compatible should be False for Cerebras."""
+    llm_core = importlib.import_module("src.llm_core")
+    assert llm_core._is_self_hosted_openai_compatible("https://api.cerebras.ai/v1") is False
+
+
+def test_apply_local_cache_affinity_skips_cerebras():
+    """_apply_local_cache_affinity must not add session_id/cache_prompt for Cerebras."""
+    llm_core = importlib.import_module("src.llm_core")
+    payload = {"messages": []}
+    llm_core._apply_local_cache_affinity(payload, "https://api.cerebras.ai/v1", "test-session-123")
+    assert "session_id" not in payload, "session_id leaked into Cerebras payload"
+    assert "cache_prompt" not in payload, "cache_prompt leaked into Cerebras payload"