diff --git a/glmocr/config.py b/glmocr/config.py
index 38fa92c..a121a40 100644
--- a/glmocr/config.py
+++ b/glmocr/config.py
@@ -430,7 +430,6 @@ def from_env(
         _KW_MAP = {
             "api_key": "pipeline.maas.api_key",
             "api_url": "pipeline.maas.api_url",
-            "model": "pipeline.maas.model",
             "mode": "pipeline.maas.enabled",
             "timeout": "pipeline.maas.request_timeout",
             "enable_layout": "pipeline.enable_layout",
@@ -442,6 +441,15 @@ def from_env(
             "cuda_visible_devices": "pipeline.layout.cuda_visible_devices",
             "layout_device": "pipeline.layout.device",
         }
+
+        # `model` is shared by both MaaS and self-hosted modes.
+        # Keep MaaS behavior while also forwarding it to OCR API so that
+        # `GlmOcr(mode="selfhosted", model="...")` works as expected.
+        if "model" in overrides and overrides["model"] is not None:
+            model_value = str(overrides["model"])
+            _set_nested(data, "pipeline.maas.model", model_value)
+            _set_nested(data, "pipeline.ocr_api.model", model_value)
+
         for kw, dotted in _KW_MAP.items():
             if kw in overrides and overrides[kw] is not None:
                 raw = overrides[kw]
diff --git a/glmocr/tests/test_unit.py b/glmocr/tests/test_unit.py
index 1771dc6..3f0d5af 100644
--- a/glmocr/tests/test_unit.py
+++ b/glmocr/tests/test_unit.py
@@ -1353,6 +1353,24 @@ def test_explicit_selfhosted_mode(self, monkeypatch):
             assert parser._use_maas is False
             parser.close()
 
+    def test_selfhosted_model_kwarg_is_forwarded_to_ocr_api(self, monkeypatch):
+        """model=... should configure self-hosted OCR request model."""
+        from glmocr.config import _ENV_MAP, ENV_PREFIX
+
+        for suffix in _ENV_MAP:
+            monkeypatch.delenv(f"{ENV_PREFIX}{suffix}", raising=False)
+        monkeypatch.setattr("glmocr.config._find_dotenv", lambda: None)
+
+        with patch("glmocr.pipeline.Pipeline") as mock_pipeline:
+            mock_pipeline.return_value.start = MagicMock()
+            mock_pipeline.return_value.enable_layout = False
+            from glmocr.api import GlmOcr
+
+            parser = GlmOcr(mode="selfhosted", model="glm-ocr")
+            assert parser._use_maas is False
+            assert parser.config_model.pipeline.ocr_api.model == "glm-ocr"
+            parser.close()
+
 
 class TestOCRClientOllamaConfig:
     """Tests for OCRClient initialization with Ollama api_mode."""