diff --git a/src/modelscope_hub/_download.py b/src/modelscope_hub/_download.py
index d1fa1b9..30e0609 100644
--- a/src/modelscope_hub/_download.py
+++ b/src/modelscope_hub/_download.py
@@ -904,13 +904,28 @@ def _lock_path(
         cache_dir: Path | None = None,
         file_path: str | None = None,
     ) -> Path:
-        """Compute the lock file path for a given repo or file."""
+        """Compute a fixed-length lock file path for a repo or file.
+
+        File systems commonly cap a single path component at 255 bytes. Older
+        versions embedded the full ``file_path`` in the lock basename, so a
+        perfectly valid remote filename could make the local lock filename too
+        long before the download even started. Keep the lock *key* semantics
+        (repo type + repo id + optional file path), but store only a stable
+        SHA-256 digest in the basename.
+        """
         base = cache_dir or self._config.cache_dir
-        safe_id = repo_id.replace("/", "___")
-        if file_path is not None:
-            safe_file = file_path.replace("/", "___").replace(".", "_")
-            return base / ".lock" / f"{repo_type}_{safe_id}_{safe_file}.lock"
-        return base / ".lock" / f"{repo_type}_{safe_id}.lock"
+        scope = "file" if file_path is not None else "repo"
+        key = "\0".join(
+            (
+                "modelscope-hub-download-lock-v2",
+                scope,
+                str(repo_type),
+                repo_id,
+                file_path or "",
+            )
+        )
+        digest = hashlib.sha256(key.encode("utf-8")).hexdigest()
+        return base / ".lock" / f"{scope}_{digest}.lock"
 
     def _download_with_resume(
         self,
diff --git a/tests/test_download_lock_path.py b/tests/test_download_lock_path.py
new file mode 100644
index 0000000..1d70d0e
--- /dev/null
+++ b/tests/test_download_lock_path.py
@@ -0,0 +1,113 @@
+"""Regression tests for download lock filename safety.
+
+Remote repositories can contain perfectly valid path components close to the
+common 255-byte filesystem basename limit. The local lock file must therefore
+never embed the raw remote path in its own basename; otherwise the lock fails
+before the actual target file is even opened.
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+from modelscope_hub._download import DownloadManager, _optional_file_lock
+from modelscope_hub.api import HubApi
+from modelscope_hub.compat.file_download import dataset_file_download
+
+_LONG_IMAGE_NAME = "a" * 215 + ".jpg"
+_LONG_FILE_PATH = f"images/{_LONG_IMAGE_NAME}"
+
+
+def _download_manager() -> DownloadManager:
+    return HubApi().downloader
+
+
+class TestDownloadLockPath:
+    def test_file_lock_name_is_fixed_length_for_long_remote_paths(self, tmp_path):
+        dm = _download_manager()
+
+        lock_path = dm._lock_path(
+            "OmniDocBench/OmniDocBench",
+            "dataset",
+            cache_dir=tmp_path,
+            file_path=_LONG_FILE_PATH,
+        )
+
+        assert lock_path.parent == tmp_path / ".lock"
+        assert lock_path.name.startswith("file_")
+        assert lock_path.name.endswith(".lock")
+        assert len(lock_path.name.encode("utf-8")) < 255
+        assert _LONG_IMAGE_NAME not in lock_path.name
+
+    def test_repo_lock_name_is_fixed_length_for_long_repo_ids(self, tmp_path):
+        dm = _download_manager()
+        long_repo_id = f"{'owner' * 30}/{'repo' * 40}"
+
+        lock_path = dm._lock_path(long_repo_id, "model", cache_dir=tmp_path)
+
+        assert lock_path.name.startswith("repo_")
+        assert lock_path.name.endswith(".lock")
+        assert len(lock_path.name.encode("utf-8")) < 255
+        assert "owner" not in lock_path.name
+
+    def test_lock_key_is_deterministic_and_file_specific(self, tmp_path):
+        dm = _download_manager()
+
+        first = dm._lock_path("owner/repo", "model", cache_dir=tmp_path, file_path="config.json")
+        second = dm._lock_path("owner/repo", "model", cache_dir=tmp_path, file_path="config.json")
+        other_file = dm._lock_path("owner/repo", "model", cache_dir=tmp_path, file_path="tokenizer.json")
+        repo_lock = dm._lock_path("owner/repo", "model", cache_dir=tmp_path)
+
+        assert first == second
+        assert first != other_file
+        assert first != repo_lock
+
+    def test_optional_file_lock_can_acquire_hashed_long_path(self, tmp_path):
+        dm = _download_manager()
+        lock_path = dm._lock_path(
+            "OmniDocBench/OmniDocBench",
+            "dataset",
+            cache_dir=tmp_path,
+            file_path=_LONG_FILE_PATH,
+        )
+
+        with _optional_file_lock(lock_path):
+            assert lock_path.exists()
+
+
+class TestCompatDownloadUsesSafeLockNames:
+    def test_dataset_file_download_accepts_long_file_path(self, tmp_path, monkeypatch):
+        """The legacy compat entry point reaches the same safe lock path.
+
+        No network is needed: patch the low-level transfer after the lock has
+        been acquired, and write the expected target file directly.
+        """
+
+        def fake_download_with_resume(
+            self,
+            repo_id: str,
+            repo_type: str,
+            file_path: str,
+            revision: str,
+            target: Path,
+            **kwargs,
+        ) -> Path:
+            target.parent.mkdir(parents=True, exist_ok=True)
+            target.write_bytes(b"image-bytes")
+            return target
+
+        monkeypatch.setattr(DownloadManager, "_download_with_resume", fake_download_with_resume)
+
+        cache_dir = tmp_path / "cache"
+        result = dataset_file_download(
+            "OmniDocBench/OmniDocBench",
+            _LONG_FILE_PATH,
+            cache_dir=str(cache_dir),
+            local_dir=str(tmp_path / "local"),
+            endpoint="https://modelscope.cn",
+        )
+
+        assert Path(result).read_bytes() == b"image-bytes"
+        lock_files = list((cache_dir / ".lock").glob("*.lock"))
+        assert lock_files
+        assert all(len(path.name.encode("utf-8")) < 255 for path in lock_files)