We're open-sourcing Hy-MT1.5-1.8B-1.25bit — a 440MB translation model that runs fully offline on your phone, supports 33 languages, and outperforms Google Translate.
At 1.8B parameters, it matches commercial translation APIs and 235B-scale models on standard benchmarks. By quantizing to 1.25-bit, memory drops from 3.3GB (FP16) to 440MB — 25% smaller and ~10% faster than prior 1.67-bit approaches, with no accuracy loss.
Covers 33 languages, 5 dialects, and 1,056 translation directions including minority languages like Tibetan and Mongolian.
Our translation model has won 30 first-place rankings in international MT competitions and is already deployed across multiple Tencent products.🏆
📲Demo APK (Android): https://t.co/DbcwLBe6vw
🤗Hugging Face:: https://t.co/NgXKeNZz4L
🔗GitHub: https://t.co/b6DqGKDelf
📄Paper:
quantization — Shrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
Why it matters
Sub-2-bit quantizationShrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.Full definition → holding benchmarkA standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.Full definition → parity at 440MB shows a useful specialist model fits on-device with no inferenceRunning a trained model to get answers — the phase where AI is actually used, as opposed to trained.Full definition → bill and no network round trip, and the method is published.
