Xiaomi released the MiMo-V2.6 weights under an MIT license on 21 September. The flagship MiMo-V2.6-Pro is a sparse MoE with 1.02T total parameters, roughly 42B of which activate per forward pass. The company also disclosed a figure that model vendors usually keep to themselves — the reinforcement learning phase cost about $3.47 million.
Key takeaways
- MiMo-V2.6-Pro: 1.02T total parameters, 42B activated, 1M-token context
- MiMo-V2.6-Flash: 309B total parameters, 15B activated, MIT license on both
- RL phase: roughly $2.62M for Pro and $850,000 for Flash
- 30 RL steps in under six days, about 750,000 trajectories per model
- 46.32 points on the Artificial Analysis Intelligence Index v4.3
One RL run instead of four
Rather than separate RL runs for code, vision and tool calling, the MiMo team mixed every task into a single session — the model card calls it "You Only RL Once". Training ran on fully asynchronous GRPO, and a single step looked like this:
One RL step, by the numbers:
A reward that separates good from merely correct
Binary pass/fail cannot tell apart two solutions that both work. Xiaomi added a grading agent that compares trajectories within each group: Groupwise Reward Synthesis builds quality rubrics from contrasting rollouts?Rollout: A single run of the model through a task from start to finish — one attempt that can then be scored., and Groupwise Advantage Redistribution shifts reward away from weaker passing answers toward stronger ones. The grader alone consumed 12.7% of Pro’s compute budget.
Strong on agents, weaker on hard code
On agentic work the model draws level with the closed flagships. The gap only shows where raw code quality matters.
| Benchmark | MiMo-V2.6-Pro | Claude Opus 5 |
|---|---|---|
| Terminal Bench 2.1 | 89.9 | 89.1 |
| AutomationBench v1.0.6 | 53.1 | 50.3 |
| ProgramBench | 26.5 | 37.0 |
| Terminal Bench 4.0 | 34.9 | 49.0 |
On the Artificial Analysis index, Pro sits ahead of DeepSeek V4.1.
Why it matters
Open weights are standard now. An open training bill is not. By publishing the cost, its breakdown and the RL environment code, Xiaomi shifts the argument from "how many parameters" to "how much compute per task interaction". It is also a test of the claim that scaling RL keeps paying off where scaling pretraining slows. Thirty steps is too short a run to settle that, but long enough for others to repeat.
What's next
- The results can be checked independently — Xiaomi shipped the technical report, training environments and RL code alongside the weights
- API pricing stayed at V2.5 levels: $0.435/$0.87 per million tokens for Pro and $0.14/$0.28 for Flash
Sources
- Hugging Face — MiMo-V2.6-Pro-RL model card
- Hugging Face — MiMo-V2.6-Flash-RL model card
- 36Kr — 罗福莉押注大规模RL、小米最强开源模型亮相
- VentureBeat — 'Better than DeepSeek': Xiaomi's MiMo-V2.6-Pro debuts as the top open weights model in the world
