Tencent Drops Pocket-Sized Hy-MT2-1.8B For 33 Language Translations

Hy-MT2-1.8B is a new open-source translation model from Tencent that handles 33 languages and can follow detailed instructions. It belongs to a family of fast-thinking translators built for real-world text like subtitles, business documents, and structured data. When compressed with its AngelSlim quantization, the entire model shrinks to 440 MB and runs 1.5 times faster.

Tencent released this 1.8B model alongside larger 7B and 30B versions, but the smallest variant specifically targets users who want capable translation on a single consumer GPU. The team also open-sourced IFMTBench, a benchmark for testing how well models follow translation instructions, and a Translator Skill for easy integration. In their multi-dimensional evaluations, the lightweight model outperformed commercial APIs from Microsoft and Doubao in overall quality.

Real-world translation with instruction control

Privacy-conscious professionals can translate sensitive documents locally without cloud dependency. Small agencies get a free tool that handles multilingual content with precise style control. Hobbyists with a single GPU can run and customize the model using the included Translator Skill and vLLM.

Performance and deployment notes

Despite its small size, Hy-MT2-1.8B often matches or beats large commercial services on translation benchmarks. The model lacks a default system prompt, so users must give explicit language and task instructions for reliable output. Tencent provides training guides for full fine-tuning and LoRA, making it straightforward to adapt the model to specialized vocabulary.

"Hy-MT2 is a family of 'fast-thinking' multilingual translation models designed for complex real-world scenarios." — Source: Hugging Face