MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement
MiMo-V2.6 scales reinforcement learning across code, visual, and agent tasks and open-sources the RL stack.
The MiMo-V2.6 report introduces an omni-modal model family trained by scaling reinforcement-learning compute after multimodal mid-training on a hybrid sliding-window-attention architecture. Asynchronous training consumes 1,568 samples and 2.7-3.7 billion tokens per step at contexts up to 1 million tokens, across code, general, visual, and cyber environments. The authors freeze the mixture-of-experts router, defend against reward hacking, and use groupwise agentic grading for long-horizon rewards. They open-source training dynamics, RL environments, and the RL framework; the abstract does not report public benchmark scores or a weights release.
- Async RL uses 1,568 samples and 2.7-3.7B tokens per step.
- Context lengths reach up to 1 million tokens.
- Environments span code, general, visual, and cyber tasks.
- The MoE router is frozen and reward-hacking defenses are added.
- Training dynamics, RL environments, and the RL framework are open-sourced.
- AI models
- MiMo-V2.6
From the original article
1 min read at huggingface.co
Reinforcement learning (RL) is the central training paradigm for advancing large foundation models towards self-improvement. This report introduces the MiMo-V2. 6 series, an omni-modal family that pushes the frontier of model intelligence by scaling RL compute.
By Core Team, Zongming Qiao, Ziyue Hua · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.