Xiaomi's MiMo-V2.6-Pro tops the open-weights index under an MIT licence — and ships the RL code, which is the part that matters
TL;DR: Xiaomi released MiMo-V2.6-Pro on 21-22 September 2026 — a sparse mixture-of-experts model with 1.02 trillion total parameters, 42 billion active, a 1M-token context window, and native text, image, video and audio input. It scores 46 on Artificial Analysis’ Intelligence Index v4.3.2, the highest open-weights result on that board, ahead of GLM-5.3 (45) and Kimi K3 (44) and level with Grok 4.7 at xHigh effort (46), which shipped the day before. Pricing is $0.435 per million input tokens and $0.87 per million output — against Grok 4.7’s $2/$6. The licence is MIT. Alongside the weights on Hugging Face and ModelScope, Xiaomi published the technical report, the RL training environments and the training code. The frontier is still above it: Claude Fable 5.1 and GPT-6 Astra sit at 53. This lands one day after Alibaba moved Qwen-Image-2.1 from Apache 2.0 to a research-only licence.
What shipped
MiMo-V2.6-Pro is a sparse mixture-of-experts model with 1.02 trillion total parameters and 42 billion activated per token. The model card describes a 681-million-parameter vision encoder, dual audio encoders totalling 435 million parameters, a five-layer speculative decoder for multi-token prediction, and a one-million-token context window. Inputs are text, image, video and audio; output is text.
The weights are published on Hugging Face and ModelScope under an MIT licence. Deployment configurations are documented for SGLang and vLLM. Artificial Analysis lists the release date as 21 September 2026; Xiaomi’s own announcement of the V2.6 series is dated 22 September.
Published pricing, per million tokens, is $0.0036 for cache-hit input, $0.435 for cache-miss input, and $0.87 for output — unchanged from the V2.5 series.
Reported evaluation results from the model card:
| Benchmark | MiMo-V2.6-Pro |
|---|---|
| DeepSWE v1.1 | 71.9 |
| Terminal-Bench 2.1 | 89.9 |
| Toolathlon-Verified | 76.9 |
| OSWorld-Verified | 82.0 |
And the independent composite, from Artificial Analysis’ Intelligence Index v4.3.2 — a blend of ten evaluations including AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode and Humanity’s Last Exam:
| Model | Intelligence Index | Weights |
|---|---|---|
| Claude Fable 5.1 (max) | 53 | closed |
| GPT-6 Astra (max) | 53 | closed |
| Claude Opus 5 (max) | 51 | closed |
| Muse Spark 1.3 (max) | 48 | open |
| GPT-5.6 Sol (max) | 47 | closed |
| MiMo-V2.6-Pro | 46 | open (MIT) |
| Grok 4.7 (xHigh) | 46 | closed |
| GLM-5.3 | 45 | open |
| Kimi K3 | 44 | open |
| Gemini 3.8 Flash | 41 | closed |
| DeepSeek V4.1 Flash | 39 | open |
Artificial Analysis puts the cost per Intelligence Index task at $0.13, output speed at roughly 125 tokens per second, and time to first token at about 2.3 seconds.
The tie is the least interesting number on the page
MiMo-V2.6-Pro and Grok 4.7 both score 46. They were released within a day of each other. One costs $6 per million output tokens and cannot be downloaded; the other costs $0.87 and can be run on hardware its user owns. That comparison writes itself, and it is worth roughly one sentence, because composite index ties are fragile in three separate ways.
The first is version drift. That 46 is Intelligence Index v4.3.2. Secondary coverage of this leaderboard routinely mixes scores from different index versions — figures of 53, 58.9 and 63 all circulate for the same top-tier models depending on which revision and which reporting scale is being quoted. A score is only comparable to another score computed by the same version of the same blend, and that qualifier is dropped more often than it is stated.
The second is effort settings. Grok 4.7’s 46 is measured at xHigh effort, one notch below the Max setting where the frontier models post their numbers, and the Grok 4.7 launch made effort labelling a live issue in its own right by reporting its own generation at xHigh against its predecessor at High. A reasoning-effort dial is a spend dial. Two models at the same index score at different effort settings are not at the same price.
The third is that benchmark families fork. Xiaomi’s model card reports Terminal-Bench 2.1 at 89.9. The Grok 4.7 launch page reports Terminal-Bench 4.0 at 38.0. Those are not a 51-point gap; they are two different evaluations that share a name, and any comparison built by putting them in adjacent cells is meaningless. The same caution applies to the DeepSWE figures circulating for this release: the model card’s 71.9 and a widely-quoted 72.57 are both attached to MiMo-V2.6-Pro in different write-ups, which is the ordinary noise of a release day and a reason to read the card rather than the coverage.
What survives all three caveats is a coarser claim, and it is enough: the best open-weights model is now within striking distance of current proprietary flagships, and seven index points below the top of the board. That gap was much wider a year ago. It has been closing steadily through Kimi K3, GLM-5.3 and Tencent’s HY4 preview, and it is not news that it closed a little further this week.
The recipe is the release
The part of this release that is genuinely unusual is what shipped alongside the weights.
Xiaomi published the full technical report, the reinforcement-learning training environments, and the RL code, with training recipes spanning coding, cyber, general, visual and music domains. The methods named in the card are specific enough to be checked: group relative policy optimisation run on very large batches — 1,568 prompts by 16 rollouts per step — plus a groupwise agentic grading scheme for self-improvement and a multi-prefix, multi-teacher on-policy distillation pass after RL.
This is a different kind of artefact from open weights. Weights are a finished product that can be run. Environments and training code are a production line that can be re-run, audited, forked and improved. The distinction has a direct commercial consequence: a lab that receives weights can serve a model, while a lab that receives the recipe can build the next one — and can check whether the published result was produced the way the report says it was.
It also puts a floor under a claim that would otherwise be unfalsifiable. The figure attached to this release across social coverage is roughly $3 million, derived by an outside reader from the technical report as about 130 hours and 75 billion tokens for a $2.6 million RL run, with pretraining added on top. Xiaomi has not confirmed it, and it plainly excludes data, failed runs, staff, and the compute amortised from the V2.5 line this model extends. It should be quoted with those caveats or not at all. But because the environments and code are public, it is a claim that can be settled by someone with the hardware, rather than argued about indefinitely — which is more than can be said for any training-cost number a closed lab publishes.
The cyber entry in the training-recipe list deserves a note of its own. GLM-5.3 shipped with coding and cyber capability as a stated axis, and open-weight security capability is one of the few places where permissive licensing and capability policy genuinely collide. An MIT licence has no acceptable-use clause; that is the point of it, and it is also what makes published offensive-security training environments a different object from published benchmark scores.
Two directions in one week
Put this release next to the one that preceded it by a day.
On 20-21 September, Alibaba published Qwen-Image-2.1 — a smaller, cheaper, more capable image generator — under a research licence permitting non-commercial use only, breaking a line of Apache 2.0 releases stretching back to 2025. On 21-22 September, Xiaomi published a trillion-parameter omnimodal model under MIT with the training stack attached.
Those are not two data points on a trend. They are two directions, chosen per release, for the same underlying reason: a lab pays for open weights out of revenue it forgoes, and it re-examines the decision when a specific checkpoint becomes valuable enough to sell. Alibaba’s image model removed a production step and became worth restricting. Xiaomi does not sell frontier inference as its business, so the calculus runs the other way.
Meta’s arithmetic produced a third answer again — Muse Spark 1.3’s open-weights release slipped and arrived behind a contributor tier that prices access in data — and Muse Spark 1.3 sits two points above MiMo-V2.6-Pro on the same index at 48. The pattern is not national, not ideological and not a vendor trait. It is per-release commercial arithmetic, and the only reliable way to read it is to open the LICENSE file for the specific checkpoint being adopted.
That is also why the consolidation of open-weight distribution matters more than it looks. A permissive licence protects the copy already distributed. It does not guarantee the registry entry stays up.
What it changes for buyers
The open-weight price floor dropped again. GLM-5.3 Flash and Qwen Flash Next held that floor in August; a model at index 46 serving at $0.435 and $0.87 per million lowers it for the capability band above them. Anyone currently paying frontier rates for work that does not need frontier capability now has a concrete, MIT-licensed comparator to price against. Against Grok 4.7’s $2/$6 that is roughly a fifth of the input cost and a seventh of the output.
Cost per completed task, not cost per token. The rate card is a quarter of the argument. A cheaper model that takes more turns, retries more often, or fails a long-running agent job outright can be more expensive per finished unit of work. The $0.13 cost-per-index-task figure from Artificial Analysis is a useful starting point and a poor substitute for replaying an internal workload — the same discipline that applies to any coding tool or agent platform evaluation.
Self-hosting is an option, not a plan. 1.02 trillion resident parameters is a serving-cluster problem regardless of how few activate per token, and the realistic path for most teams is a hosted endpoint with the MIT licence held in reserve as leverage. That is still worth having. The difference between a vendor that can change its terms and a vendor whose terms a buyer could walk away from is the entire value of a second source, and it does not require ever actually migrating.
Check the modality shape before the index score. Text, image, video and audio in, text out. That is a strong document-and-agent profile and not an answer for anyone generating media, where dedicated image and video tools remain the relevant shortlist. Developers evaluating this for coding work should weight the agentic evaluations — Terminal-Bench, Toolathlon, OSWorld — over the composite, and should note which version of each was run.
Archive what you adopt. If the licence is part of the reason for the decision, keep the weights locally. September has now produced one flagship open-weight release under MIT and one flagship reversal to non-commercial terms, five days apart. The lesson of the second is that a registry is not a shelf.
The verdict
MiMo-V2.6-Pro is the strongest open-weights model currently measured, and the tie with Grok 4.7 on a composite index will be the sentence that travels. It is the wrong sentence. Index ties are version-dependent, effort-dependent and fragile, and this one will be stale within a month.
The part that will not be stale is that a trillion-parameter omnimodal model, its technical report, its reinforcement-learning environments and its training code were published together under a licence with no commercial restriction — during the same week another lab decided its best release was worth locking down. Capability parity between open and closed weights has been arriving gradually for two years. Terms parity has not, and it is the one buyers actually sign. On that axis this release moved further than the benchmark row suggests, and it moved in the direction the score cannot measure.
Update, 23 September 2026 — the price floor argument met a counterexample the next day. The closing section above argues that MiMo-V2.6-Pro at $0.435 input / $0.87 output lowers the open-weight price floor and gives buyers “a concrete, MIT-licensed comparator to price against.” One day later that comparison acquired an awkward column: OpenAI shipped GPT-6 Luna at $0.10 input and $0.50 output — roughly a quarter of MiMo-V2.6-Pro’s input rate and 57% of its output rate, on a closed model with no weights.
This does not undo the argument; it relocates it. On hosted token price alone the open-weights leader is no longer the cheapest option in its band, so the case for MIT-licensed weights has to be made on what the licence buys rather than on the rate card — portability, the option to self-host, and insurance against a hosted provider changing terms, which is exactly the framing the FAQ above uses. What it does undermine is any procurement argument that reaches for open weights primarily to save money on inference. As of 22 September that argument no longer survives contact with OpenAI’s small tier, and buyers making it should say plainly which of the two things they are actually buying.
Frequently asked questions
Is MiMo-V2.6-Pro actually as good as Grok 4.7?
On one composite index, on one day, they tie. Artificial Analysis' Intelligence Index v4.3.2 puts MiMo-V2.6-Pro at 46 and Grok 4.7 at xHigh effort at 46, and that index is a blend of ten evaluations covering professional work, tool use, coding, science and reasoning. A tie on a blend is not a tie on any particular job. The two models also publish different benchmark suites — Xiaomi's model card reports Terminal-Bench 2.1 at 89.9, while the Grok 4.7 launch page reports Terminal-Bench 4.0 at 38.0, and those are different evaluations with the same family name, not two results on one scale. The honest reading is that MiMo-V2.6-Pro is now in the same broad capability band as a current proprietary flagship, and that the band is wide. Neither number tells a team whether the model can do its specific work, which is still something a replayed internal workload decides.
What does the MIT licence actually allow?
Commercial use, modification, redistribution, and self-hosting, with essentially one obligation: keep the copyright and licence notice. There is no non-commercial clause, no field-of-use restriction, no acceptable-use policy incorporated by reference, no user-count threshold above which separate terms apply, and no requirement to license derivatives back. A fine-tune of these weights on proprietary data can be served commercially without asking Xiaomi anything. That is materially different from both the bespoke research licences appearing on some Chinese releases and the contributor-tier arrangements appearing on some Western ones, and it is the reason this release is a second source in a way that a permissively-benchmarked but restrictively-licensed model is not.
Did it really cost $3 million to train?
Nobody at Xiaomi has said so. The figure comes from a third-party reading of the technical report circulated on X, which estimated roughly 130 hours and 75 billion tokens for a $2.6 million reinforcement-learning run, with pretraining bringing the total to about $3 million. That is an outside estimate of one training phase, not a disclosed budget, and it excludes the things that dominate a real frontier programme: data acquisition, failed runs, staff, and the compute amortised across the V2.5 line this model builds on. The number is worth knowing because it is being quoted widely, and worth discounting for the same reason. What can be verified is the recipe, because Xiaomi published it — anyone with the hardware can check the cost claim rather than argue about it.
Can a normal company actually run a 1.02T-parameter model?
Not on a workstation, and that is the practical limit on what 'open weights' buys here. The model activates 42 billion parameters per token, which governs inference speed, but the full 1.02 trillion parameters must be resident across the serving cluster, which governs how much hardware has to be bought before the first token is generated. The model card documents SGLang and vLLM deployment, so the software path is ordinary; the capital is not. For most teams the realistic route is a hosted endpoint at the published rate of $0.435 per million input tokens and $0.87 per million output, with the MIT licence functioning as an insurance policy — the option to move providers, or to bring it in-house, if the hosted terms change. That option is worth something precisely because it never has to be exercised to have value.
Why does publishing the RL code matter more than the benchmark score?
Because a score is a claim and a recipe is a capability. Open weights let a team run someone else's finished artefact; open training environments and RL code let a team build the next one, or verify that the published one was built the way the report says. Xiaomi released the technical report, the training environments and the RL code across coding, cyber, general, visual and music domains. The immediate effect is that the methods described — group relative policy optimisation over very large batches, groupwise agentic grading, on-policy distillation from multiple teachers — stop being a paper and become something other groups can reproduce and improve. The second-order effect is on price: every lab that can reproduce this result is a lab that will not pay frontier rates for work in this band.
What should a buyer do about this in the next fortnight?
Three things. Re-run a representative internal workload against MiMo-V2.6-Pro at a hosted endpoint and compare total cost per completed task against the incumbent, not cost per million tokens — the token rate is roughly a seventh of Grok 4.7's output price, and whether that survives contact with real task lengths is an empirical question. Second, archive the weights locally if the MIT licence is part of the reason for choosing it; a permissive licence protects the copy that has been distributed, not the registry listing. Third, keep the omnimodal claim in its own column: text, image, video and audio in, text out, is a specific shape, and teams that need audio or video generation are not served by it no matter how the index reads.
Sources
- Hugging Face — XiaomiMiMo/MiMo-V2.6-Pro-RL model card
- Xiaomi MiMo — official site
- Artificial Analysis — MiMo-V2.6-Pro model page
- Artificial Analysis — LLM leaderboard
- GitHub — XiaomiMiMo/verl (RL code and environments)
- Unite.ai — Xiaomi's new flagship model leads open-weight rankings with a score of 46
- Latent Space — AINews: Xiaomi MiMo-V2.6-Pro 1T-A42B
Related tool reviews
Questions or corrections? Email Pick Right. Want the full list? See all news.