Tencent Releases Hy-MT2 Translation Models: 1.8B, 7B, and 30B-A3B Support 33 Languages

Tencent has open-sourced Hy-MT2, a family of "fast-thinking" multilingual translation models designed for complex real-world scenarios. The release includes three model sizes: 1.8B, 7B, and 30B-A3B (Mixture of Experts), all supporting translation among 33 languages.

Model Specifications

All three Hy-MT2 models can follow translation instructions in multiple languages. For on-device deployment, Tencent's AngelSlim 1.25-bit extreme quantization reduces the 1.8B model's storage requirement to 440MB and increases inference speed by 1.5x.

The models are released with multiple quantization options:

• Full precision models
• FP8 quantized versions
• GGUF format for llama.cpp
• 2-bit GGUF quantization
• 1.25-bit GGUF quantization (1.8B only)

Performance Claims

According to Tencent, multi-dimensional evaluations show the models deliver strong performance across general, real-world business, domain-specific, and instruction-following translation tasks. The company claims the 7B and 30B-A3B models outperform open-source models including DeepSeek-V4-Pro and Kimi K2.6 in fast-thinking mode. Tencent also claims the 1.8B model surpasses commercial APIs from Microsoft and Doubao.

Tencent recommends temperature 0.7, top_p 0.6, top_k 20, and repetition_penalty 1.05 for the 1.8B and 7B models. The 30B-A3B model uses temperature 0.7, top_p 1.0, top_k -1, and repetition_penalty 1.0.

Benchmark Release

Alongside the models, Tencent open-sourced IFMTBench, a new benchmark for evaluating translation instruction-following capabilities. The models support various translation scenarios including terminology-aware translation, style-specific translation, personalized translation, delimiter preservation, and structured data translation.

Deployment

The models are compatible with transformers (version 5.6.0+), vLLM, SGLang, and llama.cpp. Tencent notes the GGUF format depends on their STQ kernel, released in llama.cpp PR #22836.

Tencent is partnering with WMT26 for the Video Subtitle Translation Task and offering special awards for participants using Hy-MT models in the General Machine Translation Task.

What This Means

Tencent's release adds specialized translation models to the open-source ecosystem, addressing a specific use case often handled by general-purpose LLMs. The 1.8B model's 440MB footprint after extreme quantization makes it viable for mobile and edge deployment. However, the company's performance claims comparing against commercial APIs require independent verification. The 33-language support and instruction-following capabilities suggest these models could compete with translation-specific services, though real-world performance in production environments remains to be tested.