Tencent Open-Sources Hy-MT2 Translation Models, 1.8B Shrinks to 440MB After Quantization
aigclink · x · 2026-10-09
Tencent Hunyuan has open-sourced Hy-MT2, a family of "fast-thinking" multilingual translation models in three sizes: 1.8B, 7B, and 30B-A3B (MoE), covering translation across 33 languages.
Built for on-device deployment: with AngelSlim 1.25-bit extreme quantization, the 1.8B model takes only 440MB of storage and gains 1.5x inference speed. Tencent claims the lightweight 1.8B already outperforms Microsoft and Doubao translation APIs overall, while the 7B and 30B-A3B beat open-source rivals like DeepSeek.
Beyond vanilla MT, Hy-MT2 emphasizes controllable translation and instruction following: consistent terminology, style specification, preserving delimiters, translating only user-visible text in JSON and other structured data, and context-aware translation. A good fit for MT, subtitle translation, and localization workflows.
Related event: Tencent Open-Sources Hy-MT2 Multilingual Translation Models(2 posts)→
More from Models
- OpenAI releases 372 math proof claims, including 23 Erdős problems spanning 1268 pages — burny_tech · 2026-10-09
- dhh prefers Codex as main coding model with Claude secondary, praises Sol 6.1 — npew · 2026-10-09
- Thomson Reuters built its own AI model for ~$450K, pairing it with frontier models in production — bigdata · 2026-10-09
- Step 5 Preview model appears on OpenRouter rankings, now live to try — PrajwalTomar_ · 2026-10-09
- Step 5 Preview spec sheet: 600B/27B MoE, 1M context, vision, finance reasoning — JaynitMakwana · 2026-10-09
- The Step 5 Preview finance game forces a save/invest/spend/bet choice each round — JaynitMakwana · 2026-10-09