Researchers have introduced a novel reference-free post-training methodology to enhance multilingual machine translation (MT) in open large language models. Building upon the supervised fine-tuned (SFT) MiLMMT-46-v0.1 models, the authors implemented Group Relative Policy Optimization (GRPO) to refine translation performance without relying on gold-standard reference texts. The reinforcement learning (RL) reward mechanism utilizes an average of two reference-free quality estimation models, which is gated by a language identification component to ensure linguistic accuracy. To optimize the final model, the researchers employed linear interpolation between the SFT and RL checkpoints, resulting in the MiLMMT-46-v1.0 series. Experimental evaluations across 46 languages demonstrate that this approach consistently improves translation quality compared to the initial SFT versions and outperforms existing open-source baselines such as Seed-X, HY-MT2, and TranslateGemma. Notably, the MiLMMT-46-v1.0 models achieved leading reference-free scores, rivaling proprietary systems including Google Translate, Gemini 3 Pro, and GPT-5. Additionally, the study investigated on-policy distillation (OPD), finding that while OPD can reach the quality frontier established by the RL and checkpoint interpolation method, it does not surpass it. This work provides a scalable framework for improving multilingual MT capabilities in open-source LLMs using automated quality signals.