MiMo-V2.6: What Xiaomi's RL Paper Actually Shows, and What It Leaves Unverified

Most of the model papers published this month are architecture papers. The technical report behind Xiaomi MiMo-V2.6-Pro and Xiaomi MiMo-V2.6-Flash is not. Titled MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement, submitted on 21 September 2026 and attributed to LLM-Core Xiaomi, it spends almost none of its length on the sparse mixture-of-experts backbone and nearly all of it on one question: what happens when the entire post-training budget goes into a single mixed reinforcement-learning run. DeepSeek-V4.1-Flash's report, by contrast, is a memory document — its length goes to compressed sparse attention, cross-layer KV reuse and FP4 storage. Put the two side by side and they are arguments about where capability comes from, and they do not agree.

The report is worth reading on its own terms, but it is worth reading carefully, because it is a self-report from the lab that ran the run. It names its mechanisms, shows its ablations, publishes its failures, and in the same breath reports numbers that cannot be checked from outside. Separating those two things is most of the work. This piece does that, and it is written on 23 September 2026 — two days after the checkpoints went up, while the paper is the newest and least-corroborated thing about them.

Why the date on the paper matters more than usual

Xiaomi's MiMo-V2.6-Pro and MiMo-V2.6-Flash checkpoints landed on the model hub on 21 September 2026, MIT-licensed and ungated. The report is dated the same day and reached the top of the trending list on the preprint site where it was posted. That timing is the reason this is a news piece rather than a retrospective: the paper is not a later explanation of a model everyone has already benchmarked. It is arriving alongside the weights, before anyone outside Xiaomi has published an independent run.

That ordering is also the paper's main weakness as evidence. Everything downstream of it — the DeepSWE curve, the pass-rate gains, the reward-hacking defence — is a claim about a training process that happened once, inside one lab, on one cluster, and that nobody else has reproduced. The thread that flagged the report treats that as the interesting part: this is a data-and-experiments paper, not an architecture paper, and experiments are exactly the kind of result that either replicates or does not.

Three axes of scaling, and only one of them is a hardware problem

The report's framing is that reinforcement-learning compute was scaled along three axes at once, and the third is the one that changes how you think about the method.

• Batch and throughput — 1,568 samples per update, expanded to sixteen rollouts each, so roughly 25,000 trajectories per step, consuming 2.7 to 3.7 billion tokens per step at context lengths up to one million tokens. The rollout architecture is fully asynchronous, which is what makes that batch size survivable at all.

• Environments — a single mixed run across code, general agent, visual and cyber tasks, under several agent harnesses at once, rather than separate per-domain RL runs. Xiaomi's own shorthand for this is "You Only RL Once." The argument for mixing is that gains in one capability reinforce the others; the risk is that a slow task type gets starved, which the report says it handles with adaptive scheduling.

• Grader compute — the axis that is not about GPUs in the usual sense. Binary pass/fail cannot rank two solutions that both pass, so the reward signal itself becomes the thing you scale. An agentic grader compares the sixteen attempts for a task jointly and produces a graded signal instead of a bit.

That third axis is where the phrase "self-improvement" in the title earns its place, and where the report is most exposed. A model grading its own rollouts is a reward-hacking surface, and the report treats it as one.

The two mechanisms worth actually understanding

Groupwise Advantage Redistribution

After each step, the policy produces sixteen attempts per task and all of them are placed in a shared workspace so that an evaluator model can examine them together rather than one at a time. Passing patches are then scored on five axes: whether the approach fits the problem, whether the implementation is precise without needless fallbacks, whether the change is minimal, whether it edits anything out of scope, and whether it matches the surrounding code style. Lower-ranked passes receive less positive reinforcement. A patch confirmed as a hack is reset to zero and treated as a failure.

The consequence is a training signal that pushes toward shorter, cheaper solutions rather than merely correct ones — the report frames this as steering toward fewer tokens per task. That is the part to hold onto, because the headline results later in the paper point the other way.

Groupwise Reward Synthesis

The offline half of the same idea. Rather than deriving quality purely from a live grader, the team precomputes task-specific rubrics and multiplies the binary test reward by quality and behaviour scores drawn from those rubrics. The report describes this as building the rubrics from contrasting rollouts — that is, from cases where the outcome differed, which is where the information is.

Grading is not free. The report puts grader cost at roughly 12.7% of MiMo-V2.6-Pro's total reinforcement-learning cost. That single figure is the most useful thing in the paper for anyone considering copying the method, because it converts "scale your graders" from an idea into a line item.

The frozen router is the result most teams will copy first

The report's stability section contains a finding that stands on its own, independent of MiMo-V2.6 and independent of how well the model performs.

With trainable mixture-of-experts routers, expert load drifted badly over twenty steps. The coefficient of variation across experts rose from 0.78 to 2.0. Peak load on the busiest expert went from six times the average to sixteen times. The share of cold experts — those receiving almost no traffic — went from 0.5% to 22%. Restoring the router weights to their starting values at step 20 recovered balance immediately.

Xiaomi's response was to freeze the MoE router for the run. The reasoning is not subtle: at this batch scale, routing instability is a training-dynamics problem you can simply decline to have, and the router is not where the reinforcement learning is doing its work. Whether freezing costs anything in final quality is not answered by an ablation in the published material, and that is a fair thing to want.

What the finding does not say is anything about serving. Router load balance during a training run and expert utilisation under production traffic are different questions, and the report only addresses the first.

The numbers, with their labels attached

The report's headline results are clean in shape and messy in detail, and the mess is not a scandal — it is three different measurements of three different objects that share a name.

• DeepSWE v1.1, held out — MiMo-V2.6-Pro rose from 58.4 to 72.6 across the run; MiMo-V2.6-Flash from 48.7 to 65.7. The benchmark is 113 long-horizon repository-engineering tasks judged by functional verifiers across five programming languages, and the metric is average@3 — the average verifier pass rate across three sampled attempts, which is not the same as whether any one of three attempts succeeded. Reading avg@3 as pass@3 will overstate every number on the page.

• The same benchmark, measured elsewhere — the released model cards print 71.9 for Pro and 67.9 for Flash. Xiaomi's public training dashboard printed 72.57 and 65.68. So there are three published figures per model, two of them from Xiaomi, and the direction of the discrepancy is different for each sibling. None of them is wrong; they describe a dashboard snapshot, a card entry and a report line, at different points and possibly under different harnesses.

• Distillation — MiMo-V2.6-Distill-Qwen-9B, fine-tuned from Qwen3.5-9B on MiMo-generated demonstrations, moved SWE-bench Verified from 61.1 to 66.2. This is the most transferable number in the paper, because a 9B student is a size most teams can actually run.

• An internal cybersecurity mini-benchmark — 31.3 to 47.0. Internal means internal: the tasks are not published, so the delta cannot be reproduced even in principle.

• The Artificial Analysis Intelligence Index — 46 for MiMo-V2.6-Pro. This one is different in kind. It was produced by Artificial Analysis running its own harness against the released checkpoint, not by Xiaomi, which makes it the only headline figure in this release that a reader can treat as measured rather than reported. It is also the number to compare against GLM-5.3, Kimi K3 and the closed frontier, and it sits five points under Claude Opus 5.

The report is honest about the limits of its own table, and this is the sentence worth quoting back at anyone who is not: the comparisons mix public and internal benchmarks and do not isolate the contribution of reinforcement learning from architecture or pre-training. A model that is better after RL is not proof that RL is why.

There is a second caveat, and it is the one that cuts against the framing. The reported gains generally accompany rising token use. The report presents that as sustained capability improvement rather than as efficiency, and it is right to. But GAR was explicitly designed to steer toward shorter paths and fewer tokens per task, and the aggregate result does not show the workload getting cheaper. Capability went up; cost per task did not come down. If you read "self-improvement" as "the model will need less of my money next quarter," the paper does not support that reading.

What the report leaves unverified

As of 23 September 2026, no third party has published an independent rerun of the DeepSWE, Terminal Bench or CyberGym columns. The distinction that matters is between the index point and everything else: 46 is a measurement someone else made, and 72.6 is a claim about a training run nobody can re-execute, because the run cost a reported $2.6 million for Pro and $0.9 million for Flash and is not going to happen twice.

What Xiaomi did publish that most labs do not is the training dynamics, the reinforcement-learning framework and the task environments — more than 7,000 graded environments across software engineering, vulnerability reproduction, knowledge work and web design. That is the artifact with the longest shelf life. The weights tell you what the run produced; the environments tell you how to run the experiment yourself, which is the only way any of the RL claims ever get settled.

The reward-hacking defence deserves a specific caveat. It is described as multi-layered — reward design, adversarial evaluation, anomaly detection and cross-checking between verifiers — and it is described entirely by the party that would be embarrassed by a failure of it. A defence against a model gaming a model-based grader is not the kind of thing a self-report can close out. The mechanism is named and the failure mode is acknowledged, which is more than most papers do, and it is still an open question.

What you can actually do with this today

Both checkpoints are downloadable, ungated, under an MIT licence tag: Xiaomi MiMo-V2.6-Pro-RL and Xiaomi MiMo-V2.6-Flash-RL, omnimodal, one-million-token context, with the Flash index reporting 172.9 GB of weight data across 65 shards in FP8. Xiaomi has not published a per-token rate card for the V2.6 generation, so the choice today is self-hosting against a hosted model you already pay for.

That comparison is where the paper's real-world value sits, and it is unflattering to the paper in a useful way. The claim under test is not "is MiMo-V2.6-Pro good" — the 46 answers that. It is "does reinforcement learning on mixed agentic tasks produce reasoning that transfers to my workload," and the only honest way to answer that is to run both and compare, which is a routing problem before it is a research problem. DeepSeek-V4.1-Flash is one API key away at $0.15 and $0.60 per million tokens, Qwen3.8-Flash and GLM-5.3-Flash sit on the same key, and if you want to put a self-hosted MiMo checkpoint behind the same endpoint for a week and see whether the DeepSWE curve shows up in your own evals, that is a configuration change rather than a migration. OrcaRouter does not host Xiaomi's models, and nothing here should be read as claiming otherwise — but the models you would benchmark them against are all reachable through one OrcaRouter API key at provider list price with 0% markup passed through, with automatic failover if a checkpoint you are trialling turns out to be unstable under load.

The unproven-model case is the one failover was built for. A checkpoint published two days ago, with a vendor-run evaluation and no independent rerun, is exactly the thing you want to try without putting a production path behind it.

Who should read the paper

If you run post-training, read it for the router finding and the grader-cost figure. Both are portable, both are cheap to test, and neither depends on believing anything about MiMo-V2.6's benchmark table. The frozen-router result in particular is the kind of thing that costs an afternoon to try and saves a run.

If you are choosing a model, read the index point and skip the rest. Artificial Analysis measured 46 on the released artifact; that is the one number in this release that did not come from the vendor, and it places MiMo-V2.6-Pro at the top of the open-weights field and five points behind Claude Opus 5. Everything else in the report describes how Xiaomi got there, and how Xiaomi got there is not something you can buy.

And if you are reading the coverage rather than the paper, the thing to watch for is the word "self-improvement." The report's own results show capability rising alongside token use, which is a real and repeatable finding about scaling reinforcement learning. It is not a claim that the model got cheaper to run, and the papers that follow this one are the ones that will tell you whether it eventually does.