Xiaomi has published a technical report on the MiMo-V2.6 series, an omni-modal family of foundation models whose development centers on scaling reinforcement learning (RL) compute as a path toward model self-improvement. The report frames RL as the central training paradigm for pushing large foundation models further.
Before RL begins, the team conducts mid-training on a broad multimodal corpus intended to provide ample exploration space, and builds infrastructure on a pretrained hybrid-SWA architecture to support later scale-up.
The report describes scaling RL compute along three dimensions:
- Larger batches and higher throughput: asynchronous training consumes 1,568 samples and 2.7–3.7B tokens per step at context lengths of up to 1M.
- More diverse and complex environments: spanning code, general, visual, and cyber domains under a mixture of agent harnesses.
- More grader compute: groupwise agentic grading yields more accurate reward signals for long-horizon tasks and steers the model toward shorter, more token-efficient solutions.
To keep training stable at scale, the team freezes the MoE router and establishes a multi-layer defense against reward hacking. It also builds infrastructure for mixed-task agentic RL, including a unified trajectory representation, high-concurrency multi-framework rollout, decoupled control and data planes, and training-inference consistency.
The company says it is open-sourcing the training dynamics, RL environments, and RL framework to facilitate reproduction and further research on scaled RL and model self-improvement. Model weights and a blog post accompany the report.