Skip to content

Back to the catalog

MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement
huggingface.copaper

MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement

Technical report for Xiaomi's MiMo-V2.6 omni-modal model family, which scales RL compute along three axes: asynchronous training with 1,568 samples and 2.7-3.7B tokens per step at up to 1M context, more diverse environments (code, general, visual, cyber) under mixed agent harnesses, and groupwise agentic grading that gives better reward signals and favors shorter, token-efficient solutions. Stability measures include a frozen MoE router and a multi-layer defense against reward hacking, built on a mid-trained hybrid-SWA architecture. The authors open-source training dynamics, RL environments and the RL framework, which makes it a useful reference for reproducing scaled agentic RL.

TL;DR Xiaomi's MiMo-V2.6 report on scaling agentic RL (async, 1M context, groupwise agentic grading, frozen MoE router, anti-reward-hacking) with open-sourced environments and framework.
Open original ↗
Document

What it claims

MiMo-V2.6 is Xiaomi's omni-modal model series. The report says it advances model capability mainly by scaling RL compute after a multimodal mid-training stage on a pretrained hybrid-SWA architecture. The abstract adds one thing beyond the summary: an infrastructure layer built specifically for mixed-task agentic RL, which the authors present as what makes the scale-up possible. They also say they release the training dynamics, RL environments and RL framework.

How it works

The abstract names four infrastructure components but does not describe them in detail:

  • Unified trajectory representation across task types: code, general, visual and cyber environments, run under a mixture of agent harnesses.
  • High-concurrency, multi-framework rollout. Rollouts come from more than one agent framework, not a single harness.
  • Decoupled control and data planes in the RL system.
  • Training-inference consistency handling, meaning the rollout engine and the trainer are kept aligned.

The scaling axes are the ones the summary already gives: asynchronous training at 1,568 samples and 2.7–3.7B tokens per step with up to 1M context, broader environments, and groupwise agentic grading. The stability measures are also the same: a frozen MoE router and a multi-layer reward-hacking defense. Mid-training on a broad multimodal corpus comes first and is meant to "provide ample exploration space" for RL.

How to use it

Weights and the project blog are linked from the page's community post:

The page does not list model names, parameter counts or a licence. Check the collection before you plan to adopt it. No repository URL is given for the open-sourced RL framework or environments.

Limits

  • The page is abstract-only. It reports no benchmark results, no baselines, no ablations of the frozen router or the grading scheme, and no compute totals beyond the per-step figures.
  • How the reward-hacking defense, the groupwise grader and the four infrastructure components work is only named, not explained.
  • The "cyber" environment domain is mentioned without scope or safeguards.
  • On the paper page, Hugging Face shows 0 models and 0 datasets linked to it, even though the community post points to a weights collection. Confirm that the collection actually corresponds to this report.
  • The model sizes, licence and the location of the released RL framework all have to come from the blog or the full report.

Organisations named here

  • Xiaomi
  • Hugging Face

Technologies named here

  • MiMo-V2.6 RL framework
  • MiMo-V2.6
  • MiMo-V2.6 RL environments
  • Hugging Face
  • Mixture of Experts

Techniques named here

  • Scaled asynchronous RL
  • Groupwise agentic grading
  • Mid-training on multimodal corpus
  • Frozen MoE router
  • Multi-layer defense against reward hacking
  • Mixed-task agentic RL infrastructure

Related in training