TL;DR
Xiaomi’s MiMo-V2.6 report details large-scale mixed-task agentic RL: long-horizon batches, diverse environments and harnesses, and agentic grading beyond binary test rewards. It reports training-cost curves, router-stability experiments, infrastructure failures, and a subsequent multi-teacher distillation stage.
Why it matters
This report makes several practical RL-scaling choices inspectable: how prompt groups become billion-token batches, where rollout and grading costs go, and why reward quality and expert-load stability matter alongside more compute. It also separates a mixed-RL phase from later capability integration through distillation.
Key findings
- 01
MiMo-V2.6-Pro has 1.02T total and 42B active parameters; Flash has 310B total and 15B active parameters. Both use a hybrid sliding-window/global-attention MoE backbone with multimodal encoders and agent-focused mid-training before RL.
- 02
The training setup specifies 1,568 prompts per batch with 16 rollouts per prompt, about 25K trajectories and 2.7–3.7B training tokens per step. GRPO uses asynchronous partial rollouts with staleness 4, prompt-mean loss aggregation, and the Muown optimizer; 1,568 is the prompt count, not the total trajectory count.
- 03
The authors report RL post-training spend of $2.6M for Pro and $0.9M for Flash. Pro’s reported split is 43.8% rollout, 43.5% training, and 12.7% grading. These are reported RL-phase costs, not the total cost of developing either model.
- 04
Across the RL run, DeepSWE v1.1 average@3 rises from 58.4 to 72.6 for Pro and from 48.7 to 65.7 for Flash: gains of 14.2 and 17.0 score points. Average@3 is an average over attempts, not success on at least one of three attempts. The report separately observes that broader performance gains generally accompany greater token use.
- 05
Groupwise Reward Synthesis multiplies test success by solution and behavior rubric scores for selected high-pass-rate tasks. Groupwise Advantage Redistribution grades mixed-outcome groups online, corrects confirmed hacks, and reallocates positive advantage among passing solutions; failed trajectories are not simply promoted by a quality score.
- 06
A router-freezing comparison tracks Pro’s decoder layer 9 over the first 20 steps. With a trainable router, peak expert load grows from about 6× to 16× the mean; freezing keeps it near 5.5× while benchmark performance continues to improve. This supports a stability intervention for the tested configuration.
- 07
After mixed RL, MOPD2 combines domain teachers using both full student rollouts and single-turn student continuations from teacher or SFT history prefixes. The report’s final benchmark table follows this additional stage and should not be treated as the isolated effect of RL.
Evidence & limits
The report combines architecture, mid-training, RL, grading, harness diversity, and subsequent distillation. Final comparisons do not isolate the causal contribution of each component or establish a universal compute-scaling law.
Evaluation mixes public benchmarks with internal MiMo Code, Cyber, and Visual Coding benchmarks. CyberGym environments are modified to correct flaws, so protocol differences matter when comparing outside results.
The RL learning-curve endpoints and final model table are distinct measurements: for example, the reported Pro DeepSWE RL endpoint is 72.6, while the final table lists 71.9. They should not be silently substituted for one another.
Groupwise grading can curb undesirable behavior in the reported code-only ablation, but this does not establish shorter responses across the full run: the report shows performance and total token use often rising together.
Router-load evidence is from a specified layer and paired runs, not a broad multi-seed guarantee. The report also documents network, GPU, rollout-cache, expert-imbalance, and host-memory interruptions during scaling.
The reported RL spend excludes the full pretraining and development budget. The open release claims and self-improvement framing do not establish reproduction of the complete Pro training run or autonomous recursive self-improvement.
Scaling dimensions
Models, methods & benchmarks
- Models
- MiMo-V2.6-Pro, MiMo-V2.6-Flash
- Algorithms
- GRPO, Groupwise Reward Synthesis, Groupwise Advantage Redistribution, MOPD2, MoE router freezing
- Benchmarks
- DeepSWE v1.1, AutomationBench v1.0.6, MiMo Visual Coding, ProgramBench, CyberGym
Topics