Tencent's open-source mobile-side model targets offline translation scenarios.

```

Author | Huang Yu

Tencent is accelerating the application of its large models in real-world scenarios, and this time it is focused on translation.

On April 29th, Tencent Hunyuan released and open-sourced the ultimate quantized and compressed version of the translation model Hy-MT1.5-1.8B-1.25bit, compressing a large translation model supporting 33 languages to just 440MB. No internet connection is needed; simply download and run locally on your phone.

In addition to open-sourcing the model weights, Tencent Hunyuan also launched a practical “Tencent Hunyuan Translation Demo Version,” specifically optimized for “background word extraction mode.”

Whether viewing emails locally or browsing web pages, Hunyuan Translation is always available on demand.

After experiencing the Tencent Hunyuan Translation Demo Version, Wall Street Insights found that this version also requires no network and no subscription. Currently, domestic users can download it from the Model Community, but it only runs on Android, with no iOS adaptation yet.

Tencent stated that the demo version processes everything locally, with no collection or upload of personal information, and is usable permanently after a single download.

If its actual translation capabilities rival mainstream tools like Youdao Translator and Google Translate, and it maintains permanent free usage, offline local operation, and does not upload private data, then Tencent’s move will undoubtedly shake up the competitive landscape of the translation industry.

This offline translation model from Tencent Hunyuan is built on the Hunyuan translation large model Hy-mt1.5. Hy-mt1.5 is a professional translation large model developed by the Tencent Hunyuan team, with 1.8B parameters, natively supporting 33 languages, 5 dialects/ethnic languages, and 1056 translation directions.

Tencent reveals that Hy-mt1.5 achieves translation performance comparable to commercial translation APIs and 235B-level large models. In rigorous evaluation benchmarks, its translation quality surpasses mainstream systems such as Google Translate.

In real-world business scenarios, the Hunyuan team discovered that the original 1.8B model, even at FP16 precision, still occupies 3.3GB of memory, which is too large and slow for mobile devices, making quantization and compression necessary.

Therefore, for different mobile users, Tencent has specifically launched two extreme quantization compression solutions: 2-bit and 1.25-bit.

As AI applications are booming, all parties are striving to find suitable niche scenarios for their large models.

Apart from Tencent Hunyuan, domestic and international tech giants have long been competing in the field of offline AI translation on mobile devices, with many rolling out lightweight, open-source translation models that support offline operation on devices, including Alibaba, Google, Meta, and others.

Risk Warning and DisclaimerThe market involves risks, and investment should be made cautiously. This article does not constitute personal investment advice, nor does it take into account the individual user's special investment goals, financial situation, or needs. Users should consider whether any opinions, views, or conclusions in this article suit their particular circumstances. If investing based on this, responsibility lies with the investor. ```