Chrome built-in Translator API: support, languages, speed and accuracy

The Translator API ships an on-device translation model inside Chrome. Pages call it directly from JavaScript, no cloud round trip, no API key, no per-request cost. This page tracks where the API is available right now, which language pairs the bundled model covers, and what is known about its quality and latency. The grid below is the operational answer: which browsers and platforms can run Translator.create({ sourceLanguage, targetLanguage }) in production today.

Availability matrix

Browser and platform support for the W3C-tracked Translator API as of 2026-05-25. The API surface is the same across Chromium browsers, but the underlying model and rollout status differ. Mobile remains the largest gap.

Language coverage

Chrome's bundled translation model supports 39 languages, listed in the official documentation. Pairs are downloaded on demand the first time a page requests them. Availability is checked with Translator.availability({ sourceLanguage, targetLanguage }), which returns 'available', 'downloadable', 'downloading', or 'unavailable'. Chrome deliberately hides per-pair download progress to prevent fingerprinting.

English is the hub of the bundled pack set. Non-English to non-English pairs often work, but Chrome may internally route through English and the produced translation can degrade in quality compared to a direct cloud service. There is no public list of which direct pairs ship as single models versus pivot pairs.

Accuracy

Google describes the Translator API model as an "expert model trained to generate high-quality translations" rather than a general Gemini Nano output. It is purpose-built for translation and sized to ship inside the browser, not to match a frontier cloud translator. Google has not published BLEU, COMET, or MQM scores for it.

Independent third-party benchmarks comparing the Chrome Translator API against Google Translate, DeepL, or GPT-class translators are not published as of this review date. The Chrome team itself notes that small-model translation metrics are noisy without curated reference sets and has avoided publishing comparative BLEU scores. Treat the model as competitive with mid-tier free cloud translators on the supported pairs and as substantially weaker for long, ambiguous, or domain-specific text.

Known quality caveats:

• Short inputs (single words, two-word fragments) are unreliable in the companion Language Detector API and weaker in translation context.
• Non-English to non-English pairs may pivot through English internally, compounding errors.
• There are no formality, domain, or glossary controls. The API takes a string in and returns a string out.
• Streaming output is supported, but partial chunks may revise as more context arrives.
• The model is updated on Chrome's release cadence. Same prompt, different Chrome version, possibly different output.

Speed

The API runs locally, so there is no network latency once the language pack is on disk. Google has not published per-call latency targets. In practice the first call to a new pair triggers a download (hundreds of megabytes for the base runtime, smaller per-pair packs) and blocks until the model is ready. The documentation warns that translations are processed sequentially: subsequent translations are blocked until earlier ones complete.

Performance guidance from the official docs:

• Chunk large documents and translate in parallel only by spawning multiple Translator instances when needed.
• Use the streaming API (translateStreaming) for long inputs so the UI can render as tokens arrive.
• Always show a loading indicator on first use of a pair, since the pack download can take minutes on slow connections.
• Reuse a single Translator instance for repeated calls on the same language pair to avoid repeated setup cost.

Hardware and platform requirements

These are the documented minimums for the on-device model to be available, taken from the Chrome built-in AI hardware page. Devices below the bar will report 'unavailable' even on supported Chrome versions.

Methodology

Statuses are derived from primary sources only: the Chrome for Developers documentation, the Chrome Platform Status feature entry, and the Microsoft Edge developer documentation. A row is marked stable once the API ships in the corresponding browser's stable channel without an origin trial or flag requirement. experimental means the API exists but ships behind a flag, an origin trial, or in a pre-stable channel only. unavailable means no implementation has been announced or shipped.

Cross-browser positions are tracked via Chrome Platform Status and the W3C Web Machine Learning Working Group incubation surface. Firefox and Safari operate their own translation features (Firefox runs Bergamot models locally, Safari uses Apple's system translation) but neither implements the Translator API surface, so pages cannot rely on a shared interface across all four engines.

Refresh cadence: this tracker is re-verified against the linked sources at least quarterly and whenever a Chromium stable bump touches the API. Corrections welcome via the contact link in the site footer.

Related content

• Chrome Translator API documentation
• Chrome client-side translation guide
• Chrome Language Detector API
• Chrome Platform Status: Translator API
• Microsoft Edge Translator API documentation