diff --git a/mtplx/default_models.py b/mtplx/default_models.py
index a0d5e02..877a973 100644
--- a/mtplx/default_models.py
+++ b/mtplx/default_models.py
@@ -209,6 +209,16 @@ def _public_model_id_from_metadata(path: Path) -> str | None:
             if inferred:
                 return inferred
 
+    # The config.json quantization layout (Q4 weights with a Q8 head, flat Q8,
+    # etc.) is shared by many MLX builds, so it can only distinguish the
+    # Qwen3.6-27B MTPLX Speed/Quality split — not the model family. Skip this
+    # refinement for folders that don't already identify as a Qwen3.6-27B MTPLX
+    # artifact; otherwise a third-party model like
+    # `Qwen3.6-35B-A3B-4bit-MTPLX-Optimized-Speed` would silently be served as
+    # `mtplx-qwen36-27b-optimized-quality`.
+    if _public_model_id_from_name(str(path)) is None:
+        return None
+
     config = _read_json(path / "config.json")
     quantization = config.get("quantization") or config.get("quantization_config")
     if isinstance(quantization, dict):
diff --git a/tests/test_default_models.py b/tests/test_default_models.py
index 1b73191..33f5ca5 100644
--- a/tests/test_default_models.py
+++ b/tests/test_default_models.py
@@ -294,3 +294,28 @@ def test_public_model_id_for_ref_maps_unknown_local_name_to_sanitized_id():
         public_model_id_for_ref("/tmp/My Custom Local Model!")
         == "my-custom-local-model"
     )
+
+
+def test_public_model_id_for_ref_third_party_q4_q8_keeps_folder_name(tmp_path):
+    # Regression: a third-party Qwen3.6-35B-A3B build with the same Q4-weights/
+    # Q8-head config layout as the Qwen3.6-27B Quality artifact must not be
+    # served as `mtplx-qwen36-27b-optimized-quality`.
+    model = tmp_path / "Qwen3.6-35B-A3B-4bit-MTPLX-Optimized-Speed"
+    model.mkdir()
+    (model / "config.json").write_text(
+        json.dumps(
+            {
+                "quantization": {
+                    "bits": 4,
+                    "language_model.model.layers.0.mlp.down_proj": {"bits": 8},
+                    "language_model.model.layers.0.linear_attn.in_proj_qkv": {"bits": 8},
+                }
+            }
+        ),
+        encoding="utf-8",
+    )
+
+    assert (
+        public_model_id_for_ref(model)
+        == "qwen3.6-35b-a3b-4bit-mtplx-optimized-speed"
+    )