Private, on-device translation for Apple Silicon — 67 languages, no API key.
Translate text across 67 languages entirely on your Mac. Tiny Aya Translate runs Cohere Labs Tiny Aya Global locally with MLX — nothing you translate leaves your Mac, and no API key is required.
Note: the model weights are licensed CC-BY-NC (non-commercial only) — see License.
- Local inference — no API key required; your text stays on your machine
- 67 languages across Europe, West Asia, South Asia, Asia Pacific, and Africa
- Side-by-side translation with streaming output
- Swap and download controls
- Its own "Reading Room" theme — a warm ink-and-paper palette designed dark-first, with a light companion; it follows your system appearance and can be switched from the app's settings menu
- Up to 30,000 characters and 8K tokens per input (whichever your text reaches first — the token count includes the translation instruction), and up to 8K tokens per output
- 8-bit quantized MLX inference on Apple Silicon
- Apple Silicon Mac
- 8 GB+ RAM recommended (~4 GB during inference)
- Python 3.13+ (uv installs a matching interpreter from
.python-versionif one isn't on your PATH) - uv
git clone https://github.com/darylalim/tiny-aya-translate.git
cd tiny-aya-translate
uv syncuv run streamlit run streamlit_app.pyThe app opens at http://localhost:8510 — the port is pinned in .streamlit/config.toml, so if something else already holds it the launch stops with an error rather than quietly moving to the next port. The page loads immediately; the first Translate click downloads tiny-aya-global (~3.6 GB) from Hugging Face and loads it into memory, so that first translation takes a while and later ones are instant. Each later start of the app repeats one small part of that on its first Translate: an anonymous request to huggingface.co asking whether the model is still current — it carries nothing you translate and no login token — followed by a re-download only if the model has been updated upstream. To tune the model or sampling parameters, edit the constants at the top of streamlit_app.py. The app ships with its own "Reading Room" theme — a warm ink-and-paper palette designed dark-first, with a light companion; it follows your system appearance and can be switched via the app's settings menu. Both palettes, and the reasoning behind every value, live in .streamlit/config.toml; edit that file to restyle it.
uv sync # install dependencies
uv run pytest -v # run tests
uv run ruff check --fix . # lint
uv run ruff format . # format
uv run ty check # type checkCI runs the same checks — ruff format --check, ruff check, and ty check, then pytest — on every pull request and on pushes to main, via GitHub Actions on a macos-latest runner (Apple Silicon, so mlx-lm installs natively). The checks are non-mutating: ruff format --check verifies formatting instead of applying it, so unformatted code fails CI rather than being auto-fixed.
Releases are automatic. Once the checks pass on main, a version bump in pyproject.toml is tagged and a GitHub Release is published from the conventional commits since the previous tag; the Release workflow in the Actions tab does the same thing starting from a patch/minor/major choice. Neither path stops for review, so the commit subjects since the last tag are the release notes — write them accordingly. Release history and notes are on the Releases page.
The application code in this repository is licensed under the Apache License 2.0.
It loads mlx-community/tiny-aya-global-8bit-mlx — an 8-bit MLX-quantized fork of Cohere Labs Tiny Aya Global — under CC-BY-NC. The model weights are non-commercial only, so running this app commercially would violate the model license regardless of the code license above.