Confirmed in a primary source 22 Sept 2026 Models

A phone company released the highest scoring open weights model in the world.

On September 22 Xiaomi released the MiMo-V2.6 series, four models led by a trillion-parameter flagship that scores 46 on the Artificial Analysis composite index, the top of the open weights field, with a one million token context window and more than seven thousand reinforcement learning environments behind it. Xiaomi's own claim is that at the same level of intelligence its price is a twentieth to a sixtieth of overseas models, and it left API pricing unchanged from the previous series.

Two caveats matter: the event date is September 22, which fell inside yesterday's window and was not carded, so this is carded at its true date and surfaced late; and Xiaomi published no safety documentation alongside the weights.

Until now

The open weights frontier belonged to companies whose business is models: DeepSeek, Alibaba's Qwen team, Moonshot, Z.AI and a Shanghai state laboratory. Xiaomi was a hardware manufacturer with a model project attached.

Now

Treat the top of the open weights table as a place a consumer hardware company can now reach in a single release, and treat a trillion-parameter open release shipping with no safety document as the normal case rather than the exception.

Xiaomi sells phones, televisions, rice cookers and, since 2024, electric cars. It is not a laboratory, and until this week its models were a respectable second tier. The V2.6 release puts its flagship at the top of the open weights field on an index nobody at Xiaomi controls, with a coding score that moved fourteen points in one version and a nine billion parameter distilled model scoring 66.2 on SWE-bench Verified, which is a number that would have led the whole field eighteen months ago and now runs on a workstation. The part of this that rivals cannot copy by downloading the weights is the seven thousand reinforcement learning environments, which are the training apparatus rather than the product, and Xiaomi named them in the release deliberately.

Put it beside the rest of September 22. Anthropic cut Opus pricing by 40 percent and OpenAI answered by halving two tiers within hours, both carded yesterday. Alibaba stood up in Hangzhou and published self-improvement figures no company had published before. And a phone manufacturer put a trillion-parameter model on the open web and said it performs at the level of the American flagships for between a twentieth and a sixtieth of the price. The three events are not coordinated and do not need to be: they are the same fact arriving from three directions, which is that capability at the top is no longer scarce enough to price the way it was priced in June. What did not arrive with any of it is a safety document. Treat the top of the open weights table as a place a consumer hardware company can now reach in a single release, and treat a trillion-parameter open release shipping with no safety document as the normal case rather than the exception.

How it got here

Dec 2025
Xiaomi releases mimo-v2-flash, a hybrid attention model with a 32,000 token native context.
Mar 18
mimo-v2-pro: one trillion total parameters, 42 billion active, one million token context.
Sep 22
MiMo-V2.6 series released: 46 on the Artificial Analysis index, top of the open weights field, priced unchanged from V2.5.
Sep 22
Anthropic and OpenAI cut model prices against each other; Alibaba publishes recursive self-improvement figures. Both carded in Edition 030.
Who has to react
Every lab that charges for a closed model, since the comparison Xiaomi published is against their prices and their weights cannot be downloaded; Alibaba's Qwen team, whose base model Xiaomi distilled onto and whose open weights crown this takes; DeepSeek, Moonshot and Z.AI, now measured against a hardware company's side project; the American agencies that named six Chinese firms over distillation on September 8, since a Xiaomi model distilled onto a Qwen base is the same technique used openly and legally at home; enterprise buyers who now have a free option at the top of the open table; and the safety institutes, which have a trillion-parameter model in the wild and no system card to evaluate against.

The evidence

ObservedTuesday September 22, 2026. Xiaomi released the MiMo-V2.6 series and published the release notes on its own developer site.

The whole observation, as recorded

Four models. mimo-v2.6-pro, which Xiaomi describes as our most powerful flagship reasoning model, omni-modal, ultra-high performance, trillion-parameter. mimo-v2.6-flash, a full-modality, high-intelligence, low-cost reasoning model. mimo-v2.6-pro-ultraspeed, flagship performance at up to 20 times the speed. And mimo-v2.6-distill-qwen-9b, a nine billion parameter model distilled onto Alibaba's Qwen. Figures published by Xiaomi: a score of 46 on the Artificial Analysis composite intelligence index for V2.6-Pro, which places it at the top of the open weights field; DeepSWE v1.1 rising from 48.8 to 65.7 for Flash and from 58.4 to 72.6 for Pro; SWE-bench Verified at 66.2 for the nine billion parameter distilled model; a context window of one million tokens; and more than seven thousand reinforcement learning task environments covering software engineering, vulnerability reproduction, knowledge work and web design. On design, Xiaomi says the series is comparable to Claude Opus 5 and GPT-5.6 Sol on Design Arena. Xiaomi's own pricing claim: At the same intelligence level, its price is only 1/20 to 1/60 that of overseas models. API pricing is unchanged from the V2.5 series. The predecessor, mimo-v2-pro, released March 18, 2026, was one trillion total parameters with 42 billion active and a one million token context. Xiaomi published no safety evaluation, no system card and no dangerous capability testing alongside the release; the release notes are engineering notes. The event date on Xiaomi's own page is September 22, which was inside Edition 030's window, and it was not carded there. Under the clock law it is carded here at its true date and recorded as surfacing late rather than moved forward to today.

Established

Xiaomi released the MiMo-V2.6 series on September 22, 2026, including a trillion-parameter flagship, a low-cost variant, a high-speed variant and a nine billion parameter distilled model.
Xiaomi published a score of 46 on the Artificial Analysis composite intelligence index for the flagship, DeepSWE v1.1 scores of 65.7 and 72.6, SWE-bench Verified of 66.2 for the small model, a one million token context and more than seven thousand reinforcement learning environments.
Xiaomi stated that at the same intelligence level its price is a twentieth to a sixtieth that of overseas models, and left API pricing unchanged from the V2.5 series.
No safety evaluation, system card or dangerous capability testing was published with the release.

Not established

That the benchmark figures hold up. Every score quoted here was published by Xiaomi, including its position on a third-party index, and no independent evaluation of these models has been published.
That the price claim is comparable. A twentieth to a sixtieth of an overseas price depends on which model, which tier and which token mix, none of which Xiaomi specifies.
That the weights are freely usable. The release notes do not state a licence, and Alibaba removed commercial rights from a Qwen image model two days earlier, so the terms matter and are not established.
That Xiaomi can repeat this. One release at the top of an index is not a cadence, and the company has no track record of holding a frontier position.

Why it weighs this much

Signal strength

CostHigh

CommitmentHigh

ReversibilityIrreversible

PersistenceYears

Cost is high: training a trillion-parameter model and building more than seven thousand reinforcement learning environments is a capital programme, not a project, and Xiaomi paid for it out of a hardware company's balance sheet. Commitment is high because publishing the weights forecloses selling access to those weights, which is the business model the flagship labs run, and because the pricing claim sets a public mark the company has to keep meeting. Reversibility is the strongest dial on this card and it is irreversible: downloaded weights cannot be withdrawn, patched or licensed back, which is the whole difference between this release and a price cut. Persistence is years because an open model at this level becomes the base other people build on, and those derivatives outlive the original release.

What to watch

weakens Xiaomi still leads the open weights field on the Artificial Analysis composite index on December 31, 2026.
strengthens A different organisation takes the top of the open weights index before December 31, 2026.

The words

Open weights: A model whose trained numbers are published, so anyone can download it, run it on their own machines and change it. Once published it cannot be recalled.
Parameters: The adjustable numbers inside a model. A trillion of them means very large capacity and very large hardware requirements to train, though not necessarily to run, since only some are active per query.
Artificial Analysis index: An independent firm's single composite score for model capability across a common set of tests. Xiaomi quoted 46 for its flagship; the index is third-party, the claim of leading the open field is Xiaomi's.
Reinforcement learning environment: A simulated task with a score attached, used to train a model by rewarding good attempts. Xiaomi says it built more than seven thousand of them, and unlike weights, those do not travel when someone downloads the model.

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.