unslothai/unsloth · error · SttModelNotDownloadedError

STT model '{model_id}' is not downloaded. Download it in Set

Error message

STT model '{model_id}' is not downloaded. Download it in Settings, then Voice, before loading it.

What it means

Raised by _ensure_model_downloaded() as SttModelNotDownloadedError when _cached_model_paths(model_id) returns None, meaning the mtmd (multimodal Whisper) model files are not present in the local cache. Both load() and transcribe() call it up front so a missing model fails fast before any server start or audio decode.

Source

Thrown at studio/backend/core/inference/stt_mtmd_sidecar.py:792

        self._update_in_progress = True
        try:
            with self._start_lock, self._lock:
                model_was_active = self._process_alive()
                self._release_locked()
                yield model_was_active
        finally:
            self._update_in_progress = False

    @staticmethod
    def _reserve_free_port() -> tuple[socket.socket, int]:
        sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
        sock.bind(("127.0.0.1", 0))
        return sock, sock.getsockname()[1]

    def _ensure_model_downloaded(self, model_id: str) -> tuple[str, str]:
        paths = _cached_model_paths(model_id)
        if paths is None:
            raise SttModelNotDownloadedError(
                f"STT model '{model_id}' is not downloaded. "
                "Download it in Settings, then Voice, before loading it."
            )
        return paths

    def load(
        self,
        model: Optional[str] = None,
        request_cancel_event: Optional[threading.Event] = None,
    ) -> None:
        if request_cancel_event is not None and request_cancel_event.is_set():
            raise SttTranscriptionCancelledError("Transcription cancelled.")
        self._raise_if_update_in_progress()
        model_id = resolve_mtmd_model_id(model)
        binary = ensure_engine_available()
        # Startup happens outside _lock (it is slow), so this keeps two callers
        # from each spawning a server and orphaning the first.
        with self._start_lock:

View on GitHub (pinned to 203007d190)

Solutions

Download the model in Settings > Voice before loading it.
Verify the model cache directory exists and contains the model and mmproj files for that model_id.
Call the download API/route for the model first, then retry the load.

Defensive patterns

Strategy: validation

Validate before calling

```python
from studio.backend.core.inference.stt_mtmd_sidecar import _cached_model_paths

if _cached_model_paths(model_id) is None:
    prompt_download(model_id)  # Settings > Voice flow
```

Try / catch

```python
try:
    sidecar.load(model_id)
except SttModelNotDownloadedError:
    start_download_flow(model_id)  # then retry after completion
```

Prevention

Check the model cache before offering the model in the UI.
Disable model choices whose download never completed.
Trigger downloads eagerly from Settings rather than lazily at dictation time.

When it happens

Trigger: load(model) or transcribe_bytes(...) with a model id whose GGUF model/mmproj files were never downloaded via the Settings > Voice download flow, or whose cache directory was deleted/moved.

Common situations: Fresh install with no model downloaded; user pointed the cache at a wiped directory; model files removed after a disk cleanup; requesting a second model that was never fetched.

Related errors

AI-assisted analysis of unslothai/unsloth@203007d190 (2026-08-15). Data as JSON: /api/errors/6949373d2ef9a470. Report an issue: GitHub.