Hy-MT2

Tencent's Apache 2.0 translation models in 1.8B, 7B and 30B-A3B sizes, for 33 languages, with a 440 MB on-device build.

In short. Hy-MT2 is a family of AI translation models from Tencent. You choose a size: a tiny one that fits on a phone, a mid-size one for a single GPU, and a larger one for a server. All three translate between 33 languages and can follow instructions such as keeping tags or glossary terms intact.

Summary

Tencent open-sourced Hy-MT2 on 21 May 2026, following HY-MT1.5 from December 2025. It has three sizes: 1.8B and 7B dense models, and a 30B-A3B mixture-of-experts model with about 3 billion active parameters. All of them translate between 33 languages, and the language table on the model card also lists Tibetan, Kazakh, Mongolian, Uyghur and Cantonese. They follow translation instructions written in several languages. They are "fast-thinking" models: they answer without a long reasoning phase. The 7B checkpoint is about 16 GB in BF16 and the 30B-A3B about 60 GB, or about 31 GB in FP8.

Tencent's report says the 7B and 30B-A3B models beat DeepSeek-V4-Pro and Kimi K2.6 in fast-thinking mode on its translation tests, and that the 1.8B model beats mainstream commercial translation APIs, including those of Microsoft and Doubao, overall. For phones and laptops, Tencent's AngelSlim toolkit compresses the 1.8B model to 1.25 bits, which shrinks it to 440 MB and speeds it up about 1.5 times. These are Tencent's own results. Tencent also published IFMTBench, a benchmark for instruction-following in translation.

The weights are under Apache 2.0, and GGUF builds exist for llama.cpp. The prompts matter: the model card gives templates for plain translation, terminology, context, delimiter preservation and structured data such as JSON or HTML, and the recommended sampling settings differ between the dense and the MoE models. Use transformers 5.6 or newer. The 30B-A3B model is a mixture-of-experts model; it is listed here as a variant of the same entry.

Good for

  • Self-hosted translation of documents, subtitles and app strings across 33 languages
  • On-device translation with the 1.25-bit 1.8B build
  • Translation that must keep tags, placeholders or glossary terms unchanged
  • Translation involving Tibetan, Kazakh, Mongolian, Uyghur or Cantonese, which the model card lists as supported

Watch out for

  • Quality claims versus commercial APIs and DeepSeek or Kimi models come from Tencent's own tests, so evaluate on your own text
  • It is a translation model, not a general chat or reasoning model
  • It needs transformers 5.6 or newer and trust_remote_code
  • The 1.25-bit and 2-bit builds are extreme quantizations, so test their quality on your own languages

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the Hy-MT2-1.8B variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
Hy-MT2-1.8B BF16 2.04B 4.69 GB 2.35 GB 1.41 GB Laptop or phone
Hy-MT2-7B BF16 8.03B 18 GB 9.23 GB 5.54 GB Laptop or phone
Hy-MT2-30B-A3B (MoE) BF16 30.1B 3B active 69 GB 35 GB 21 GB Gaming GPU
Hy-MT2-30B-A3B-FP8 FP8 30.1B 3B active 69 GB 35 GB 21 GB Gaming GPU

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is Hy-MT2?
Hy-MT2 is a family of AI translation models from Tencent. You choose a size: a tiny one that fits on a phone, a mid-size one for a single GPU, and a larger one for a server. All three translate between 33 languages and can follow instructions such as keeping tags or glossary terms intact.
Can I use Hy-MT2 commercially?
Yes. Hy-MT2 is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Hy-MT2 open source?
Hy-MT2 is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run Hy-MT2 on my own computer?
Yes. The smallest variant, Hy-MT2-1.8B, needs about 1.41 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was Hy-MT2 released?
Hy-MT2 was released by Tencent on May 21, 2026.
How long a context window does Hy-MT2 have?
Hy-MT2 supports a context window of up to 256K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
TranslateGemma Google DeepMind 4.3B – 27.4B 2K Gemma Terms (Conditional) 2.97 GB Laptop or phone
Index-Translate Bilibili 2.27B – 36B 3B active 256K Apache 2.0 (Permissive) 1.44 GB Laptop or phone
PLaMo 3 Translate Preferred Networks 890M – 31.8B 32K PLaMo Community (Conditional) 0.614 GB Laptop or phone
Qwen3.5 Alibaba Qwen 873M – 403B 17B active 256K Apache 2.0 (Permissive) 0.602 GB Laptop or phone