diff --git a/src/gpuma/models/fairchem.py b/src/gpuma/models/fairchem.py
index 767cb2e..b001d64 100644
--- a/src/gpuma/models/fairchem.py
+++ b/src/gpuma/models/fairchem.py
@@ -67,7 +67,6 @@ def _load_fairchem_torchsim(config: Config) -> Any:
 
     _load_hf_token_to_env(config)
     model_path = _verify_model_path(config)
-    model_name, model_cache_dir = _verify_model_name_and_cache_dir(config)
     # Fairchem internally only accepts "cuda" or "cpu"; _setup_fairchem_device
     # calls torch.cuda.set_device(N) when a specific GPU is requested so that
     # Fairchem's internal device resolution picks the correct GPU.
@@ -75,8 +74,14 @@ def _load_fairchem_torchsim(config: Config) -> Any:
     torch_device = torch.device(backend_device)
 
     if model_path:
+        # model_name is deliberately not resolved on this branch.
+        # _verify_model_name_and_cache_dir raises when model_name is empty, so
+        # calling it unconditionally rejected a config that supplies only a
+        # local checkpoint -- which _load_fairchem_calculator accepts. The same
+        # config then worked in sequential mode and failed in batch mode.
         uma_model = FairChemModel(model=model_path, task_name="omol", device=torch_device)
     else:
+        model_name, model_cache_dir = _verify_model_name_and_cache_dir(config)
         uma_model = FairChemModel(
             model=model_name,
             model_cache_dir=model_cache_dir,
diff --git a/tests/models/test_fairchem.py b/tests/models/test_fairchem.py
index 4adb073..59238f2 100644
--- a/tests/models/test_fairchem.py
+++ b/tests/models/test_fairchem.py
@@ -249,3 +249,70 @@ def test_calculation_required_tracks_the_cache(self):
         assert wrapper.calculation_required(atoms, ("energy",)) is False
         atoms.positions[0][0] += 0.1
         assert wrapper.calculation_required(atoms, ("energy",)) is True
+
+
+class TestFairchemLocalCheckpoint:
+    """A config supplying only ``model_path`` must work on both paths.
+
+    These stub out ``FairChemModel``/``pretrained_mlip`` rather than downloading
+    a checkpoint: the behaviour under test is purely which config fields each
+    loader insists on *before* it constructs anything, so the real model would
+    add a token requirement without adding coverage.
+    """
+
+    @staticmethod
+    def _checkpoint_only_config(tmp_path):
+        ckpt = tmp_path / "local.pt"
+        ckpt.write_bytes(b"")
+        return Config({
+            "model": {"model_type": "fairchem", "model_name": "", "model_path": str(ckpt)},
+            "technical": {"device": "cpu"},
+        })
+
+    def test_torchsim_accepts_checkpoint_without_model_name(self, tmp_path, monkeypatch):
+        """Batch mode used to demand model_name even with a checkpoint given.
+
+        _verify_model_name_and_cache_dir raises on an empty model_name, and it
+        was called unconditionally, so this config failed in batch mode while
+        working in sequential mode.
+        """
+        import torch_sim.models.fairchem as ts_fairchem
+
+        sentinel = object()
+        seen = {}
+
+        def fake_model(**kwargs):
+            seen.update(kwargs)
+            return sentinel
+
+        monkeypatch.setattr(ts_fairchem, "FairChemModel", fake_model)
+
+        from gpuma.models.fairchem import _load_fairchem_torchsim
+
+        assert _load_fairchem_torchsim(self._checkpoint_only_config(tmp_path)) is sentinel
+        assert seen["model"].name == "local.pt"
+
+    def test_calculator_accepts_checkpoint_without_model_name(self, tmp_path, monkeypatch):
+        """The sequential path already worked; pin it so the two stay in step."""
+        import fairchem.core as fc
+
+        sentinel = object()
+        monkeypatch.setattr(
+            fc.pretrained_mlip, "load_predict_unit", lambda **kw: sentinel
+        )
+        monkeypatch.setattr(fc, "FAIRChemCalculator", lambda **kw: sentinel)
+
+        from gpuma.models.fairchem import _load_fairchem_calculator
+
+        assert _load_fairchem_calculator(self._checkpoint_only_config(tmp_path)) is sentinel
+
+    def test_model_name_still_required_without_a_checkpoint(self):
+        """Dropping the check entirely would defer the error to a confusing place."""
+        from gpuma.models.fairchem import _load_fairchem_torchsim
+
+        config = Config({
+            "model": {"model_type": "fairchem", "model_name": ""},
+            "technical": {"device": "cpu"},
+        })
+        with pytest.raises(ValueError, match="Model name must be specified"):
+            _load_fairchem_torchsim(config)