Tencent open-sources Hy-MT2 translation models

Tencent is pushing translation models into the same release cadence and deployment stack that frontier LLMs get, which is a meaningful signal for teams that need low-latency multilingual systems rather than generic chat.

• –The 1.8B variant is tuned for edge use, with extreme quantization that makes on-device translation plausible without abandoning the model family
• –The 30B-A3B MoE model gives Tencent a high-end option for teams that want quality without paying the full dense-model inference cost
• –Shipping IFMTBench alongside the models matters because translation quality is often about instruction adherence, formatting, and terminology handling, not just BLEU-style scores
• –The release is more deployment-ready than academic, with Transformers, vLLM, and SGLang paths already documented
• –For developers, this strengthens the open multilingual translation stack and raises the bar for what “open-source MT” can look like in production