Xiaomi's MiMo-V2.6 runs full-parameter RL on a 310B model across 1,000+ TPUs
Xiaomi released MiMo v2.6, and @tokenbender quickly followed with an in-depth long thread, arguing that the real star of this release is reinforcement learning rather than any flashy components. The paper's core idea is to explicitly scale three variables: batch size and throughput, environment diversity, and grader compute. The author notes that the official justification for large batches rests mainly on throughput and GPU-scaling convenience.
Confirmed
- The MiMo v2.6 paper advocates large batches from the perspective of throughput and GPU-scaling convenience, explicitly scaling the three RL levers: batch size and throughput, environment diversity, and grader compute.
- Section 4.2.5 trains with multiple harnesses; tokenbender considers this commendable, since open-source users each have their own preferences and often build harnesses suited to their own tasks.
- Environment synthesis follows the generally good principles of data synthesis: a clear spec, sampling from a rich pool of real-world scenarios, and curating high-quality seed tasks with grounding. The author's verdict: "nothing fancy, but these principles are non-negotiable."
- The paper discusses CodeMidas's source-code-driven synthesis approach.
Unconfirmed
- The long-horizon task portion seems to be done mainly via agent-driven synthesis; tokenbender doubts its effectiveness and says he is interested in following up to verify.
Why it matters
- tokenbender points out that MiMo's tasks are not built from scratch but grounded from diverse sources, which pure self-play cannot replace; the tasks are close to everyday workflows, so the author is hopeful that MiMo v2.6 could replace Sol for daily work.
2026-09-22 ~ 2026-09-23 · 10 related posts
- Episode 1: Xiaomi MiMo RL Training Details Spark Debate: Dramatic Gains in Just 30 Steps(2026-09-22, 15 posts)
- Episode 2: Xiaomi's MiMo-V2.6 runs full-parameter RL on a 310B model across 1,000+ TPUs(2026-09-22, 10 posts)
- Episode 3: Xiaomi's MoE RL stability paper: router-aware GRPO tweaks explained(2026-09-22, 2 posts)
Primary sources
- Xiaomi MiMo v2.6 ships with RL as the hero, scaling batch size, env diversity and grader compute — tokenbender ·
- Xiaomi MiMo v2.6 ships: a deep dive into its RL scaling of batch size, environments, graders — tokenbender ·
- MiMo v2.6 details: grounded environment synthesis and multi-harness training for open source users — tokenbender ·
- Full-Parameter RL on TPUs: peano_ai Runs 310B MiMo-V2.6 Across 1,000+ TPUs — simonguozirui · 2026-09-22
- Xiaomi MiMo runs full-parameter RL on 310B model across 1000+ TPUs with JAX — AccBalanced · 2026-09-22
- [source] Xiaomi MiMo v2.6 ships with RL as the hero, scaling batch size, env diversity and grader compute — tokenbender · 2026-09-22
- Thread recap: MiMo v2.6's environment scaling principles grounded in diverse real sources — tokenbender · 2026-09-22
- MiMo leans on CodeMidas-style source-driven synthesis and agent-driven long-horizon tasks — tokenbender · 2026-09-22
- [source] MiMo v2.6 details: grounded environment synthesis and multi-harness training for open source users — tokenbender · 2026-09-22
- Dev breaks down paper: multi-harness training and environment synthesis principles — tokenbender · 2026-09-22
- Xiaomi's Luo Fuli Recaps MiMo-V2.6 RL Run: 25K Agent Trajectories Per Step — bookwormengr · 2026-09-22
- Xiaomi MiMo-V2.6 scales RL to 3.7B tokens per step, targeting self-improvement — Dr_Singularity · 2026-09-23
1 near-duplicate retellings: tokenbender