Researchers from Hugging Face and Xiaomi have released MiLMMT-46-v1.0, a family of open multilingual translation models that achieve state-of-the-art results without relying on reference translations during post-training. The work, detailed in a new paper, demonstrates how reinforcement learning (RL) can be effectively applied to improve translation quality across a wide range of languages.
The team started with the supervised fine-tuned (SFT) MiLMMT-46-v0.1 models and applied Group Relative Policy Optimization (GRPO). The reward function averages two reference-free quality estimation models and is gated by language identification, ensuring the reward is appropriate for each language. After RL, they linearly interpolated the SFT and RL checkpoints to produce the final v1.0 models.
Across 46 languages, the resulting models consistently outperform their SFT counterparts. They also surpass strong recent open baselines, including Seed-X, HY-MT2, and TranslateGemma, and achieve leading reference-free scores against evaluated proprietary systems such as Google Translate, Gemini 3 Pro, and GPT-5.
The authors also explored on-policy distillation, finding that it reaches but does not surpass the quality frontier achieved by RL with checkpoint interpolation. The models and code are released to facilitate future research, and a live demo is available on Hugging Face Spaces.