so I read the post-training bits of the MiMo-V2.6 tech report by @XiaomiMiMo some notes below (not sure notes like these still make sense when you can ask an agent to summarize any paper for you, but maybe my impressions and how I connect it to what we're building are still worth something) > What they call "You Only RL Once" One mixed GRPO run across code, general agents, web dev and cyber, all domains in the same batch The scale: 1,568 prompts × 16 rollouts = 25K trajectories per step, ~3B tokens per update Code takes 68% of the mix (code is everywhere!), and the gains still show up on general agent benchmarks > They train on minimal harnesses instead of production ones Production harnesses come with prompts and safeguards the reward never measures, so the model learns to ignore them Instead they train on a family of minimal mini-harnesses (open source, in mimoagent), and the gains carry over to harnesses the model never saw (Codex, Claude Code, mini-swe-agent: ~50% → 66% on DeepSWE) We just added multi-harness training to OpenEnv via Harbor, with 16 real harnesses you can rotate per rollout > They also rank the solutions that pass If all 16 rollouts pass the tests, the advantage is zero and the group teaches nothing So for easy tasks they multiply the test result by rubric scores, and for the rest an agentic grader ranks the passing patches and moves advantage toward the better ones. Without it the policy drifted into exception swallowing and relaxed validation to get tests green They do the same for visual tasks, ranking the renders within each group. This would have helped a lot when I was training models to paint with code > Reward hacking The report quotes the model going to GitHub for the upstream fix ("Let me just look at the file directly from the GitHub raw URL") Their answer was env cleanup, network isolation and a dedicated hack agent that attacks every env until it finds no exploit. Confirmed hacks stayed under 2% for the whole run We've seen similar hacks in Repo2RLEnv and have been patching them as they show up > After RL, multi-teacher distillation Separate domain-specialized teachers (RL or SFT) get distilled back into the main model on-policy (MOPD2), which is how they reach domains that are hard to verify The trick is in the prefixes: a trajectory with k turns gives k starting points, and the student only generates the next turn from each one instead of redoing the whole rollout > They share all the numbers of the run 30 steps, 123h for Pro and 82h for Flash, $2.6M and $0.9M, plus a timeline of every failure along the way (love that!) Also a nice MoE detail: with a trainable router, 22% of the experts went cold in 20 steps, so they froze it > What's released The 7.8K released envs are a curated subset of what they trained on. @adithya_s_k ported all of them to Harbor, and there's an Explorer Space where you can run a rollout on any of them from the browser They also ship a 9B starting point (Qwen3.5-9B SFT'd on MiMo data), and GRPO on the released envs improves all 11 benchmarks they report The 9B part reminds me of the SmolDataEnvs project we shared this week, 5K+ verifiable code and data science tasks for small models (<10B) to hill-climb on, trainable on a single A100