EXPLAINER · TUESDAY, SEPTEMBER 22, 2026

Xiaomi MiMo-V2.6 explained: how a 1T-parameter MIT-licensed model got within a few points of Opus 5 on agent work, where it still fails, what it costs to run, and the distillation question

Xiaomi's launch post gives you a benchmark table and a price. This page reads the table for you, benchmark by benchmark, explains the training method in plain words, tells you what hardware each of the three sizes actually needs, sets Anthropic's distillation case next to Xiaomi's own wording, and lists what nobody outside Xiaomi knows yet.

This explains reporting by Xiaomi MiMo, September 21, 'MiMo-V2.6: Scaling Reinforcement Learning Toward Self-Improvement'.
Read the original first: https://mimo.xiaomi.com/mimo-v2-6

In one minute

Xiaomi released MiMo-V2.6 on September 21. Three checkpoints on Hugging Face, all MIT: Pro-RL at 1.02 trillion total parameters with 42 billion active, Flash-RL at 309 billion with 15 billion active, and a 9 billion parameter distill of Qwen3.5-9B.
Pro and Flash take text, images, video and audio in one model, with a 1 million token context and a built-in speculative decoder that drafts 7 tokens per step.
On Xiaomi's own table, Pro is within about 3 points of Claude Opus 5 on DeepSWE, OSWorld, Toolathlon, GDPval and Terminal Bench 2.1, and ahead on AutomationBench. Artificial Analysis scores it 46, first of 114 open-weight models.
The same table shows it far behind on Terminal Bench 4.0 (34.9 vs 49.0 for Opus 5), ProgramBench (26.5 vs 37.0) and every security benchmark with a closed-model column (ExploitBench 47.9 vs 70 to 78.5).
The API costs $0.435 per million input tokens and $0.87 output for Pro, $0.14 and $0.28 for Flash. Grok 4.7, released the same day with the same 46 score, is $2 and $6.
The method is one giant reinforcement learning run across coding, agents, vision and cyber at once, with an AI grader that ranks passing answers against each other, followed by on-policy distillation from multiple teacher models. Xiaomi released the training framework and about 7,000 graded tasks.
Anthropic's September 11 threat report says Xiaomi ran more than 400,000 exchanges through Claude in March and April to harvest training data. Xiaomi's public training dashboard carried a line reading 'Claude Distill Requests: hidden.' Xiaomi has not responded.
Running Pro takes 16 GPUs across two nodes in Xiaomi's recipe. Flash fits a single 8-GPU node. The 9B runs on one consumer GPU. Most readers should start with the 9B or the API.

What shipped, exactly

On Monday, September 21, Xiaomi's MiMo team published three model checkpoints to Hugging Face and ModelScope, a technical report, a launch post, and a training toolkit. The models are MiMo-V2.6-Pro-RL, MiMo-V2.6-Flash-RL and MiMo-V2.6-Distill-Qwen-9B. All three carry the MIT license in their model cards. MIT is the short permissive license: you can use, copy, change, sell and ship the weights, and the only condition is that you keep the copyright notice. There is no research-only clause and no commercial-use exception, which is what set Qwen-Image-2.1 apart on Sunday.

Pro is a sparse mixture-of-experts model with 1.02 trillion total parameters, of which 42 billion are active for any given token. Mixture of experts means the model is split into many small specialist blocks, called experts, and a router picks 8 of 384 for each token. You store all trillion parameters, but you only compute with 42 billion at a time, which is why a model this size can answer at 124 tokens per second on Xiaomi's API. Flash is the same design at 309 billion total and 15 billion active.

Both take text, images, video and audio as input and produce text. The vision encoder is 681 million parameters, the audio path about 435 million. Context is 1 million tokens. Both ship with a 5-layer multi-token-prediction drafter that guesses 7 tokens ahead so the main model can verify them in one pass, which is how the speed number is reached.

The 9B model is different. It is Qwen3.5-9B fine-tuned on 77.4 billion tokens of data that the big MiMo model generated, covering code, general agent tasks, visual coding and cyber. Xiaomi calls it 'a starting point for open research in agentic reinforcement learning.' Its card shows it roughly doubling the base model on agent tasks: Terminal Bench 2.1 from 27.0 to 37.1, Toolathlon from 25.9 to 35.2, AutomationBench from 5.0 to 30.3.

Alongside the weights, Xiaomi released the full RL training framework and, per The Decoder's read of the report, about 7,000 ready-made tasks with automatic graders across software development, cybersecurity, office work and web design, plus about 1,000 music composition tasks. There is also a paid API with a fourth variant, Pro-UltraSpeed, at 20 times the output speed for 10 times the price.

Reading the benchmark table honestly

Xiaomi's model card compares Pro and Flash against Claude Opus 5, GPT-5.6 Sol and Claude Fable 5. Note who is not in the table: GPT-6 Astra and Claude Fable 5.1, the current top closed models. That choice makes the table look closer than a full comparison would. A Hacker News reader filled in the gaps for Terminal Bench 4.0 from public leaderboards: Astra 59.6, Fable 5.1 55.1, Opus 5 49.0, MiMo Pro 34.9.

Where Pro is close, per Xiaomi's numbers. DeepSWE v1.1, which measures fixing real software issues: Pro 71.9, Opus 5 74.0, Sol 73.0, Fable 5 70.0. OSWorld-Verified, computer use: 82.0 versus 83.4 and 83.0. Toolathlon-Verified, tool calling: 76.9 versus 80.6. GDPval-AA 2.1, office work scored by human raters: 1673 versus 1708. Terminal Bench 2.1: 89.9 versus 89.1, so slightly ahead. AutomationBench: 53.1 versus 50.3, ahead. Agents' Last Exam: a tie at 31.6. That is a real result. A year ago no open model was inside 10 points of the closed leaders on this kind of work.

Where Pro is not close. Terminal Bench 4.0, the newer and harder terminal suite: 34.9 versus 49.0 for Opus 5, a 14-point gap that grows to 25 against Astra. ProgramBench: 26.5 versus 37.0. And every security benchmark with a closed-model column: ExploitGym 17.8 versus 22.1, 30.3 and 28.4; ExploitBench 47.9 versus 70.0, 78.5 and 78.0; SEC Bench Pro 66.3 versus 79.1. On CyberGym and Xiaomi's own cyber bench, where Pro scores 94 and 80, there is no closed-model number to compare against.

The pattern is consistent. On tasks with a clear checkable answer and a fixed environment, the RL run got Pro to the frontier. On tasks that need long, messy, multi-step work in a terminal, or adversarial security work, the gap is still a generation wide. Flash sits 3 to 6 points under Pro on most rows, and further behind on cyber.

Artificial Analysis, which runs its own independent suite, gives Pro 46 on its Intelligence Index. That is first of 114 open-weight models, and it is exactly Grok 4.7's score from the same day. Fable 5.1 and GPT-6 score 53. So the honest one-line version is: MiMo-V2.6 Pro is a Grok 4.7 you can download, not an Opus 5 you can download.

The training method, from first principles

Most reasoning models today are trained in two stages. First the model learns to predict text. Then it gets reinforcement learning: it tries a task many times, a grader says which attempts succeeded, and the model is nudged toward the kind of attempt that succeeded. The usual grader is binary. A code fix passes its tests or it does not.

Xiaomi's report describes three changes to that recipe. The first is scale and mixing. Instead of separate RL runs for coding, for agents, for vision and for cyber, it runs one mixed batch across all of them: 1,568 prompts, each tried 16 times, per training step, with different tool harnesses mixed into the same batch. The claim is that skills transfer, so a model trained on harness A behaves well on harness B it never saw. The algorithm is a fully asynchronous version of GRPO, which stands for Group Relative Policy Optimization. In plain words, GRPO compares each attempt to the other attempts at the same task and rewards the ones that did better than their siblings, which avoids needing a separate model to estimate how hard each task is.

The second change is the grader. A binary pass or fail cannot tell you which of five passing solutions was best. Xiaomi adds an agentic grader that compares attempts within a group. Offline, Groupwise Reward Synthesis looks at contrasting attempts and writes a task-specific rubric, then blends rubric quality with the test result. Online, Groupwise Advantage Redistribution ranks the passing attempts and moves reward toward the better ones. The stated effect is that the model learns shorter paths and uses fewer tokens per task, because among two passing answers the shorter one now wins. Xiaomi frames this as a self-improvement loop: the model is judged against its own samples, so the bar rises as the model does.

The third change is what happens after RL. MOPD2 stands for Multi-Prefix Multi-Teacher On-Policy Distillation. Distillation means training a student model to reproduce a teacher's outputs. On-policy means the student generates the beginning of an answer itself and the teacher only supplies what comes next, so the student learns from its own mistakes rather than from a script. Multi-prefix means the student is also trained from saved teacher trajectories and supervised demonstrations, reusing their history so a decision point can be trained without regenerating everything before it. Multi-teacher means more than one teacher model. The report, as far as we can read it, does not name the teachers.

Two more details matter. The cold start for alignment is self-correction: the model rewrites its own misaligned turns into grounded next steps before RL begins. And Xiaomi says it hardened environments, screened for adversarial inputs and cross-checked verifiers to fight reward hacking, which is when a model finds a way to pass the grader without doing the task. Given the summer's escapes at OpenAI and Google, that paragraph is the one to read closely in the report.

What it costs to run each size

The model card ships serve commands, and they tell you the hardware. For Pro on SGLang, the recipe is tensor parallel 16 across two nodes with expert parallel 16, meaning 16 GPUs total and a fast interconnect between the two machines. The vLLM recipe uses tensor parallel 8, which implies a single 8-GPU node with enough memory per card to hold roughly 500GB of weights at 4-bit, or more at higher precision. Either way, Pro is a data-center deployment. It is not a workstation model.

Flash at 309B total is the one a well-funded team might self-host. At 8-bit it is about 310GB of weights, at 4-bit about 155GB, which fits one 8-GPU node comfortably and a 4-GPU node tightly. The active parameter count of 15 billion is what sets speed per token; the total count is what sets memory. This is the tradeoff every MoE model makes.

The 9B distill is the one most readers should touch first. Nine billion parameters is about 18GB at 16-bit and under 6GB at 4-bit. It runs on one consumer GPU or a recent laptop. The card gives a single SGLang command with the MiMo reasoning parser and a short OpenAI-compatible Python call with thinking turned on.

For everyone else there is the API, which is also on OpenRouter and has a documented Claude Code configuration. Pro is $0.435 per million input tokens on a cache miss, $0.0036 on a cache hit, and $0.87 per million output. Flash is $0.14 and $0.28. Artificial Analysis works out about $0.13 per Intelligence Index task for Pro. Grok 4.7, same day, same 46 score: $2 and $6. Those prices are the reason a working team should care about this release even if it never downloads a weight.

The distillation question, laid side by side

On September 11, Anthropic published a threat intelligence report covering December 2025 to August 2026. It named seven China-based labs it says ran large-scale distillation against Claude, about 190 million exchanges in total: Alibaba, Moonshot, DeepSeek, Zhipu, MiniMax, SenseTime and Xiaomi. The Xiaomi case is filed as GTG-16008. Anthropic's description, as reported by The Hacker News: 'more than 400,000 exchanges observed over 20 days in March and April 2026, in which Xiaomi replayed user conversations and coding sessions from its own MiMo models to Claude, through OpenClaw and OpenCode coding harnesses, to bolster training data used for future models.' Anthropic calls the practice 'illegal distillation' and says it responded by banning accounts from unsupported regions and by introducing preserved thinking, which stops new API accounts from altering what precedes Claude's reasoning.

Separately, when Xiaomi opened its live training dashboard on September 15, readers found a line item labeled 'Claude Distill Requests: hidden.' Forkast reported it on September 17. Xiaomi has not explained the line.

Now set Xiaomi's own words next to that. The model card describes MOPD2 as multi-teacher on-policy distillation and does not name the teachers. The Decoder notes the report 'offers almost no documentation of where the earlier training and teacher data for the internal distillation of teacher models came from.' None of this proves that Claude was a teacher for V2.6. Anthropic's window is March and April; the V2.6 RL run began in September, though the models that seeded it were built in between. It does mean the launch post's 'built in public' framing covers the RL run and not the data that fed it.

Why a working reader should care. If you ship a product on these weights, the license is clean and the MIT terms are yours. The question is not legal exposure for you. It is whether the capability you are buying is durable, or whether it depends on continued access to a closed model that has now closed the door. That is a supply-chain question, and the answer is not in the report.

What this changes, and for whom

For a team paying frontier API prices for agent tasks with checkable outputs, code fixes, form filling, document extraction, tool calls, there is now an open model at roughly a fifth to a twentieth of the cost that Xiaomi's numbers and Artificial Analysis's put within a few points of Opus 5. The right move is a side-by-side on your own tasks, not a switch.

For a team doing security work or long autonomous terminal sessions, nothing changes yet. The gap on those benchmarks is 15 to 30 points and the closed models keep it.

For anyone who wants to run a serious model in the building, Flash and the 9B distill are the candidates, and today's MacStories and Tim Dettmers items on this page are the hardware side of the same story.

For the industry, the Anthropic report and this release are two halves of one event. The best open model in the world just shipped, on time, with a dashboard, from a lab Anthropic named eleven days earlier. How the two facts relate is the question of the month.

Who is affected

Case	Status
Teams paying frontier prices for checkable agent work (code fixes, tool calls, extraction)	A credible cheaper option exists as of Monday. Test it on your own tasks before moving anything.
Security and red-team tooling built on frontier models	Not affected. MiMo trails the closed models by 15 to 30 points on every cyber benchmark with a comparison column.
Teams that want to self-host	Flash fits one 8-GPU node. The 9B distill fits one consumer GPU. Pro needs 16 GPUs across two nodes.
Anyone building on MiMo weights commercially	MIT license, no restrictions. Whether the capability is durable given the distillation allegation is a separate, open question.
Anthropic and other closed labs	The distillation case is now attached to a top-ranked release. Expect more account restrictions like preserved thinking.

What to do

Serve MiMo-V2.6-Distill-Qwen-9B on one GPU with the command in its model card and run ten of your real agent tasks through it next to your current model.
If you use OpenRouter or Claude Code, try MiMo-V2.6-Flash on a coding task this week; Xiaomi documents the Claude Code configuration.
Before choosing MiMo for terminal-heavy or security work, look at the Terminal Bench 4.0 and ExploitBench rows on the model card, not the headline DeepSWE number.
If you plan to fine-tune, download Xiaomi's released task set and graders; 7,000 graded agent tasks is a training asset on its own.
Read the reward-hacking paragraph in the technical report before you trust the model's self-reports in an autonomous loop.

What is still unknown

Which teacher models MOPD2 used. The report describes multi-teacher distillation and does not name the teachers, as far as we could read it.
What 'Claude Distill Requests: hidden' on the training dashboard referred to. Xiaomi has not explained it.
Whether Xiaomi disputes Anthropic's GTG-16008 account. No response has been published as of Tuesday morning.
Independent replication of Xiaomi's benchmark table. Only the Artificial Analysis score of 46 is third-party so far.
The full technical report. The PDF is hosted on Hugging Face and could not be machine-read for this page; the method description above comes from the model cards and secondary coverage of the report.
Total training cost. The Decoder cites $2.62 million for Pro and $0.85 million for Flash; Forkast's dashboard figures of $432,000 a day suggest those cover the RL phase only.
Memory figures for Flash and Pro above are estimates from parameter counts at 4-bit and 8-bit, not Xiaomi's stated numbers.

Sources

Today's full edition: AI News Report · every headline, every morning.