fxtranslate 0.4.2

Firefox Translations neural machine-translation engine — native (compiled Rust), batteries-included model management.

fxtranslate

Translate with Firefox Translations models from Python — a library powered by the same lightweight, CPU-only models Firefox ships for on-device translation. The engine is native compiled Rust (fast, CPU-only, no GPU), and it discovers, downloads, and caches the models for you; no API keys, no network round-trips to a translation service, nothing leaves your machine after the model download.

Unlike the npm/wasm build, the Python wheel ships the fast, batteries-included engine: the gemmology SIMD int8 GEMM kernel, mmap, ICU4X sentence segmentation, and built-in networked model discovery/download/caching. Prefer a native binary, a Rust dependency, or Node? See Related packages.

Install

$ pip install fxtranslate

The first release ships an sdist plus one platform wheel (built on the release machine). On other platforms pip compiles from source — that needs a Rust and C++ toolchain, and it falls back to the portable scalar kernel automatically if no SIMD backend is wired for your target — until the full binary-wheel matrix lands.

Library, batteries-included

Lead with this — the native advantage over the wasm build. One call discovers the model for the pair, downloads and hash-verifies it into a local cache, builds the engine, and is ready to translate:

from fxtranslate import Translator

t = Translator.load("en", "es")          # discover → download → cache → ready
print(t.translate("The weather is nice today."))
# El clima es agradable hoy.

# translate_long segments multi-sentence / long input and translates each part
# within the model's context window, instead of truncating.
print(t.translate_long(long_article))

Library, bring-your-own model

If you already have the model files on disk (recommended for production — see The models), pass the bytes directly. You supply the marian .bin model and the source/target SentencePiece vocabularies (the same file for most pairs; only CJK pairs differ) — pass the source vocab twice for a shared-vocab pair:

from fxtranslate import Translator

model = open("en-es/model.bin", "rb").read()
vocab = open("en-es/vocab.spm", "rb").read()
t = Translator(model, vocab, vocab)       # src vocab twice for shared-vocab pairs
print(t.translate("The weather is nice today."))

Discovery helpers

fxtranslate.discovery re-exports the pure model-discovery/routing surface the batteries-included path is built on — over a Remote Settings records body, returning native Python objects (no JSON strings) — for building your own model-management flow:

from fxtranslate import discovery

route = discovery.resolve_route(records_json, "en", "es")  # {"kind": "direct", ...}
cat = discovery.catalog(records_json, "en")           # {"bidirectional": [...], ...}
pairs = discovery.model_pairs(records_json)           # list[(src, trg)]
recs = discovery.parse_records(records_json)          # list[dict]
sents = discovery.segment_sentences(text)             # list[str] (ICU4X UAX #29)
plain = discovery.verify_and_decompress(zst_bytes, sha256_hex)  # bytes

The models

fxtranslate uses the models Firefox ships. They're discovered through Mozilla's Remote Settings and downloaded from Firefox's CDN, then cached under the platform-native cache directory — ~/Library/Caches/fxtranslate/models on macOS, $XDG_CACHE_HOME/fxtranslate/models on Linux, %LOCALAPPDATA%\fxtranslate\models on Windows — one subdirectory per language pair.

That hosting is provisioned for Firefox, not for third-party traffic, and neither the endpoints nor the availability of any particular model are a stable contract for downstream projects. Translator.load is a convenient way to try the engine; if you're shipping a product on top of it, mirror the model files you depend on, serve them from infrastructure you control, and load them with the bring-your-own-model Translator(...) constructor.

Related packages

The same engine is published across ecosystems on one shared version, so you can pick the artifact that fits:

• fxtranslate (crates.io) — the Rust engine library. Native SIMD int8 kernel, optional built-in model management.
• fxtranslate-cli (crates.io) — the native Rust CLI, installed with cargo install fxtranslate-cli.
• fxtranslate (npm) — the Node CLI + library (WebAssembly core).

License

MPL-2.0. The engine is a Rust port of the Firefox Translations inference engine.