<2-3 sentences summarising what the project addresses and solves, the target audience, and the benefit.> Hy-MT2 is a family of "fast-thinking" multilingual translation models designed for complex, real-world scenarios. It targets developers and researchers needing efficient, high-performance translation across 33 languages, offering models from 1.8B to 30B parameters with significant on-device optimization and instruction-following capabilities, aiming to outperform existing open-source and commercial solutions.

How It Works

This family includes 1.8B, 7B, and 30B-A3B (MoE) models. A key feature is AngelSlim, a 1.25-bit extreme quantization that reduces the 1.8B model to 440MB and boosts inference speed by 1.5x for on-device deployment. The models are engineered for complex real-world tasks and robust instruction following, with performance claims benchmarked against leading alternatives.

Quick Start & Requirements

• Primary install/run command: Hugging Face transformers (pip install transformers>=5.6.0). GGUF versions for llama.cpp and source builds for vLLM/SGLang are also available.
• Non-default prerequisites: Python 3.12 recommended for vLLM. GPU acceleration is beneficial.
• Estimated setup time or resource footprint: 1.8B model with 1.25-bit quantization requires ~440MB storage.
• Links: Hugging Face (https://huggingface.co/collections/tencent/hy-mt2), ModelScope (https://modelscope.cn/collections/Tencent-Hunyuan/Hy-MT2), Report (https://arxiv.org/abs/2605.22064).

Highlighted Details

• 7B and 30B-A3B models outperform DeepSeek-V4-Pro and Kimi K2.6 in fast-thinking mode.
• 1.8B model surpasses mainstream commercial APIs (Microsoft, Doubao) overall.
• AngelSlim 1.25-bit quantization enables 440MB footprint and 1.5x speedup for the 1.8B model.
• Introduces IFMTBench for evaluating translation instruction-following.
• Partnered with WMT26 for "Video Subtitle Translation Task" and "General Machine Translation Task."

Maintenance & Community

• Contact: hunyuan_opensource@tencent.com.
• Platforms: Models available on Hugging Face and ModelScope.
• Activity: Recent releases in May 2026 (Hy-MT2), Dec 2025 (HY-MT1.5), Sep 2025 (Hunyuan-MT).

Licensing & Compatibility

• License type: Not explicitly stated in the README, requiring clarification for adoption.
• Compatibility notes: No specific notes on commercial use or closed-source linking.

Limitations & Caveats

• Models lack a default system prompt, necessitating careful prompt engineering.
• The absence of a clear license is a significant adoption blocker.
• Detailed performance validation requires consulting the referenced report.