Tencent open-sources Hy-MT2 1.8B–30B

Tencent has open-sourced a new multilingual translation model family that it says is built for both cloud-scale and on-device use. The release, posted on May 21 across Tencent Hunyuan’s GitHub, Hugging Face collection and model site, includes three main sizes: Hy-MT2-1.8B, Hy-MT2-7B and Hy-MT2-30B-A3B. Tencent also released IFMTBench, a benchmark it said is designed to test translation instruction-following. The company describes Hy-MT2 as a “fast-thinking” translation model family for real-world multilingual work rather than a general chatbot repackaged for translation. Tencent said the models support translation among 33 languages and can follow translation-specific instructions, including structured translation, glossary-based translation, contextual translation and style-adapted translation. ### What exactly did Tencent release? (github.com) Tencent’s public release consists of the three base model sizes plus multiple deployment formats. The Hugging Face collection shows standard checkpoints, FP8 versions and GGUF variants, including low-bit files for the 1.8B model aimed at local deployment. The GitHub repository also includes a technical report, training and inference code, and the IFMTBench benchmark. (github.com) Tencent said in the repository news log that it open-sourced Hy-MT2-1.8B, Hy-MT2-7B, Hy-MT2-30B-A3B and IFMTBench on May 21, 2026. ### Why is the 1.8B model getting so much attention? Tencent’s main pitch is that the smallest model is meant to be usable on phones and other constrained devices. (huggingface.co) The company said AngelSlim 1.25-bit quantization cuts the 1.8B model’s storage requirement to 440 MB and improves inference speed by 1.5 times for on-device deployment. Tencent also said that lightweight model “surpasses mainstream commercial APIs” from Microsoft Translator and ByteDance’s Doubao overall in its evaluations. (github.com) That claim appears in both the GitHub README and Tencent’s model page, though the public materials summarize the result rather than publishing a neutral third-party comparison. ### How does Tencent position the larger models? Tencent said the 7B and 30B-A3B models outperform open-source models including DeepSeek-V4-Pro and Kimi K2.6 in what it calls “fast-thinking mode.” The 30B-A3B model is a mixture-of-experts system, while the 7B model is positioned as a denser middle tier between the mobile-focused 1.8B release and the larger flagship. (github.com) The company’s model page says Hy-MT2 improves instruction-following over the previous generation and performs strongly on benchmarks including Flores200 and WMT25, as well as in domain and business translation scenarios. (github.com) Tencent attributed those gains to multilingual monolingual and bilingual data used in continued pretraining, followed by post-training with on-policy distillation and reinforcement. ### What are the limits on using it? Tencent released Hy-MT2 under a “Tencent HY Community License Agreement.” The license text says the agreement does not apply in the European Union and is limited to a defined territory, which means developers will need to review terms closely rather than treating the release as a standard permissive open-source license. That matters because “open-sourced” in AI often means weights and code are public, but usage rights can still be narrower than software developers might expect under licenses such as Apache or MIT. (aistudio.tencent.com) That is an inference from the published license terms. ### What comes next for the model family? Tencent said it is partnering with WMT26 on the “Video Subtitle Translation Task,” and said participants using Hy-MT models in both the general machine translation task and the video subtitle task can compete for special awards sponsored by Hunyuan. (github.com) Tencent also published a Hy-MT2-Translator Skill to help developers integrate the models into translation workflows. (github.com)