Tencent Hunyuan open-sources Hy-MT2 translation models and launches the "Tencent Hy Translation" mini program.

AIMPACT Update: On May 21 (UTC+8), according to monitoring by Beating, Tencent Hunyuan officially open-sourced the Hy-MT2 series of multilingual translation large models, releasing three versions with parameters of 1.8B, 7B, and 30B-A3B (MoE architecture). The series supports translation between 33 languages and five Chinese dialects. Weights and code have been simultaneously released on Hugging Face, ModelScope, and GitHub, and have been adapted and deployed across multiple hardware platforms including ARM, Qualcomm, Intel, Muxi, and TianShu ZhiXin. Additionally, Tencent Hunyuan has partnered with WMT26 to co-host the “Video Subtitle Translation Competition.” On-device deployment is a key focus of this upgrade. Leveraging Tencent’s proprietary AngelSlim/Sherry framework, the 1.8B model achieves ultra-low-precision quantization at 1.25-bit, reducing its size to just 440MB. On Apple’s A15 chip, this version delivers 1.5x faster inference compared to its previous 4-bit quantized counterpart, enabling smooth offline inference on mainstream smartphones. Furthermore, the 7B and 30B-A3B versions outperformed DeepSeek-V4-Pro on the FLORES-200 benchmark and achieved 96% to 99% of the performance level of Gemini 3.1 Pro in real-world and domain-specific evaluations. The official also open-sourced the translation instruction-following evaluation benchmark, IFMTBench. The mini-program “Tencent Hy Translation,” built on this model foundation, is now available for trial, featuring voice input, style presets, and customizable instructions. iOS and Android apps supporting offline on-device inference will be released soon. (Source: BlockBeats)