Skip to main content
Back

Tencent's Hy-MT2-7B Beats DeepSeek on Translation and Runs Locally

Tencent's Hy-MT2-7B arrives as a GGUF via Unsloth, bringing purpose-built 33-language translation that beats DeepSeek and commercial APIs to local hardware

  • Unsloth releases GGUF quants of Tencent's Hy-MT2-7B, making it runnable locally via Ollama, llama.cpp, and LM Studio.
  • Purpose-built for translation: supports 33 languages with instruction-following for terminology, style, delimiters, and structured data like JSON.
  • Beats bigger models: 7B and 30B-A3B outperform DeepSeek-V4-Pro and Kimi K2.6 in fast-thinking mode; 1.8B beats Microsoft Translator and Doubao APIs.
  • Extreme compression: the 1.8B variant quantizes to just 440 MB via AngelSlim 1.25-bit with a 1.5x inference speedup, phone-deployable.
  • New benchmark released: IFMTBench is open-sourced alongside the model, with 7,344 samples covering single and multi-constraint translation tasks.
  • Apache 2.0 license, available now on Hugging Face and ModelScope; llama.cpp users need to build from source for the STQ kernel.

Unsloth's GGUF quantization of Tencent's Hy-MT2-7B is now live on Hugging Face, making one of the most capable open-source translation models to date trivially runnable on consumer hardware. The base model comes from Tencent's Hunyuan team, and the story behind it is worth understanding before you reach for ollama run.

Not your average multilingual LLM

Most open LLMs treat translation as a side feature, trained to do coding, reasoning, chatting, summarization, agents, and everything else under the sun. Hy-MT2 takes the opposite approach. It is a family of fast-thinking multilingual translation models designed for complex real-world scenarios, covering three model sizes (1.8B, 7B, and 30B-A3B MoE) that all support translation among 33 languages and follow translation instructions in multiple languages.

The phrase "fast-thinking" here has a specific meaning. Traditional LLM translation uses a "slow thinking" approach, understanding full semantics first before generating. Tencent introduced a "fast-thinking" paradigm: react like a professional human translator, cutting unnecessary reasoning overhead. The result is lower latency without sacrificing translation fidelity.

Benchmark numbers that raise eyebrows

Multi-dimensional evaluations show Hy-MT2 delivers outstanding performance across general, real-world business, domain-specific, and instruction-following translation tasks. The 7B and 30B models outperform open-source models such as DeepSeek-V4-Pro and Kimi K2.6 in fast-thinking mode, while the lightweight 1.8B model surpasses mainstream commercial APIs from providers such as Microsoft and Doubao overall.

Across benchmarks like FLORES-200, WMT25, COMETKiwi, and GEMBA, the Hy-MT2 family performs extremely well, especially the larger 30B-A3B model. On the instruction-following side, Hy-MT2-7B achieves 89.73 / 72.67 / 83.14 on IFMTBench's Simple, Complex, and Total scores respectively, outperforming Gemma4-E4B and other medium-sized open-source baselines. Hy-MT2-30B-A3B further improves to 90.20 / 75.94 / 84.69, achieving the best overall IFMTBench score among small-to-medium-sized models.