diff --git a/app/models/megadetector.py b/app/models/megadetector.py
index a66ef52..d797baa 100644
--- a/app/models/megadetector.py
+++ b/app/models/megadetector.py
@@ -9,6 +9,7 @@
 
 from PytorchWildlife.models import detection as pw_detection
 from .base_model import BaseModel
+from ..utils.checkpoint_utils import get_checkpoint_path
 
 logger = logging.getLogger(__name__)
 
@@ -55,18 +56,27 @@ def load(self, model_path: str, device: str, **kwargs) -> None:
             device: Device to load the model on (e.g., 'cpu', 'cuda', 'mps')
             **kwargs: Additional parameters including:
                 - model_id: The model variant (e.g., 'MDV6-yolov10-e')
+                - checkpoint_path: Optional explicit weight file (local path
+                  or URL, resolved like every other model type). Absent ->
+                  PytorchWildlife downloads its pinned weights itself.
                 - conf: Default confidence threshold
         """
         # Use model_id from kwargs as the version
         model_id = kwargs.get('model_id')
         if not model_id:
             raise ValueError("Model ID must be provided in the model configuration")
-        
+
+        # Resolve an explicit weight through the shared resolver so
+        # MODEL_BASE-style configs can pin/serve this model like the others,
+        # instead of always downloading from PytorchWildlife's zenodo URL.
+        local_weights = get_checkpoint_path(kwargs.get('checkpoint_path'))
+
         logger.info(f"Loading MegaDetector model {model_id} on device {device}")
-        
+
         # Load the model using model_id as version
         try:
             self.model = pw_detection.MegaDetectorV6(
+                weights=local_weights,
                 version=model_id,
                 device=device,
                 pretrained=True
diff --git a/tests/test_megadetector_checkpoint.py b/tests/test_megadetector_checkpoint.py
new file mode 100644
index 0000000..f205f8e
--- /dev/null
+++ b/tests/test_megadetector_checkpoint.py
@@ -0,0 +1,81 @@
+"""Tests for explicit checkpoint_path support in the MegaDetector loader.
+
+Every other model type resolves its weights through
+checkpoint_utils.get_checkpoint_path (local path or URL). MegaDetector
+was the exception: it always let PytorchWildlife fetch weights from its
+hardcoded zenodo URL at load time — a hidden internet download on every
+cold start that MODEL_BASE cannot control or pin.
+
+PytorchWildlife needs GPU/network extras that aren't importable in the
+test environment, so a stub module captures the constructor call.
+"""
+import sys
+import types
+
+import pytest
+
+
+@pytest.fixture
+def megadetector_model(monkeypatch):
+    """Import app.models.megadetector against a stubbed PytorchWildlife."""
+    captured = {}
+
+    class FakeMegaDetectorV6:
+        def __init__(self, weights=None, device="cpu", pretrained=True,
+                     version="yolov9c"):
+            captured.update(weights=weights, device=device,
+                            pretrained=pretrained, version=version)
+
+    detection = types.ModuleType("PytorchWildlife.models.detection")
+    detection.MegaDetectorV6 = FakeMegaDetectorV6
+    models = types.ModuleType("PytorchWildlife.models")
+    models.detection = detection
+    pw = types.ModuleType("PytorchWildlife")
+    pw.models = models
+
+    monkeypatch.setitem(sys.modules, "PytorchWildlife", pw)
+    monkeypatch.setitem(sys.modules, "PytorchWildlife.models", models)
+    monkeypatch.setitem(sys.modules, "PytorchWildlife.models.detection", detection)
+    monkeypatch.delitem(sys.modules, "app.models.megadetector", raising=False)
+
+    from app.models.megadetector import MegaDetectorModel
+    yield MegaDetectorModel, captured
+    monkeypatch.delitem(sys.modules, "app.models.megadetector", raising=False)
+
+
+def test_checkpoint_path_reaches_pytorchwildlife(megadetector_model, tmp_path):
+    MegaDetectorModel, captured = megadetector_model
+    weights = tmp_path / "MDV6-yolov10-e-1280.pt"
+    weights.write_bytes(b"w")
+
+    model = MegaDetectorModel()
+    model.load(model_path="", device="cpu", model_id="MDV6-yolov10-e",
+               checkpoint_path=str(weights))
+
+    assert captured["weights"] == str(weights)
+    assert captured["version"] == "MDV6-yolov10-e"
+
+
+def test_absent_checkpoint_path_keeps_auto_download(megadetector_model):
+    """Without checkpoint_path the current behavior is preserved:
+    PytorchWildlife receives weights=None and fetches its own."""
+    MegaDetectorModel, captured = megadetector_model
+
+    model = MegaDetectorModel()
+    model.load(model_path="", device="cpu", model_id="MDV6-yolov10-e", conf=0.1)
+
+    assert captured["weights"] is None
+    assert captured["pretrained"] is True
+
+
+def test_missing_local_checkpoint_fails_fast(megadetector_model):
+    """A configured-but-absent local weight is a config error, not a cue
+    to silently download something else."""
+    MegaDetectorModel, captured = megadetector_model
+
+    model = MegaDetectorModel()
+    with pytest.raises(Exception) as excinfo:
+        model.load(model_path="", device="cpu", model_id="MDV6-yolov10-e",
+                   checkpoint_path="/nonexistent/weights.pt")
+    assert "not found" in str(excinfo.value).lower()
+    assert captured == {}, "PytorchWildlife must not be constructed on bad config"