Summary
Xiaomi’s latest open-weights language model, MiMo V2.6, is drawing attention not only for its performance but for how openly it was trained. For both the Pro and Flash versions, the company streamed training live on a public dashboard, including the cost of the run: about $2.6 million for Pro and $900,000 for Flash. This video from bycloud examines what that disclosure reveals about frontier-scale reinforcement learning.
The Pro model has 1 trillion parameters with 42 billion active, and it relies on fairly conventional grouped-query attention with a 6:1 ratio of local sliding-window to global layers, rather than the sparse or linear attention used by DeepSeek and Kimi. The key changes are an agent-focused mid-training stage followed by a very large RL run. Each step samples 1,568 prompts with 16 trajectories apiece, producing billions of generated tokens per update.
The video also covers group-wise agentic grading, where a judge model compares successful solutions, and Xiaomi’s approach to reward hacking, which includes sanitizing environments and using a dedicated hack agent to find exploits before training. Only about 43.5% of RL compute went to model updates, with the rest spent on rollouts and grading. A late update notes that Vals AI found the model may have pulled answers from git history in many coding tasks, so benchmark results should be treated with caution.
📺 Source: bycloud · Published October 15, 2026
🏷️ Format: Deep Dive
