Xiaomi Releases MiMo-V2.6-Pro and Flash Open-Source Models

On September 21, 2026, Xiaomi MiMo released two models, MiMo-V2.6-Pro and MiMo-V2.6-Flash, with the release post showing approximately 22,800 views on X. This is the most certain layer of this event: the time, product names, and release action.

The second layer is product positioning. The two models are described as open-source, omnimodal, and advanced through scaled reinforcement learning. A distinction is needed here: the open-source and omnimodal descriptions come from the publisher and relays, belonging to source claims rather than independently verified facts.

The third layer is the results that sparked discussion. The publisher claims Pro scored 46 on the Artificial Analysis Intelligence Index, ranked 6th, and says it has become one of the top open-weight models currently available. These are the most viral numbers in this release, and also the weakest link in the evidence chain.

46 points, 6th place—these two numbers are specific enough to be easily cited as established fact. But based on existing evidence, they come from the publisher and relay posts, with no official confirmation from Artificial Analysis seen in the sources, nor any third-party independent reproduction.

More critically, the sources do not explain what sub-items make up this index, what range 46 points falls into, or under what criteria the 6th place ranking was produced. Without this context, the score and ranking can only be treated as source claims, not directly converted into capability conclusions.

Therefore, reading "6th place" as "the world's sixth-strongest model" is an inference beyond the evidence. The only verifiable statement is one sentence: the publisher claims it ranked 6th.

Xiaomi claims that MiMo-V2.6-Pro performs comparably to Claude Opus 5 and GPT-5.6 Sol on most agentic benchmarks. This is the strongest comparison statement in the release, and also the sentence with the most incomplete information.

The sources do not give the specific names or scores of these agentic benchmarks, nor do they explain which items "most" covers, how large the gaps are, or on which items it is not comparable. Without these coordinates, "comparable" cannot be tested or reproduced.

For readers, the reasonable approach is to record it as the publisher's comparison claim, not as an established capability ranking.

The two models are described as open-source, omnimodal, and advanced in a "built in public" manner. This description comes from Xiaomi MiMo and belongs to source claims.

However, the sources do not provide key specifications such as model parameter count, context length, license terms, or weight download addresses, nor do they explain the Flash version's specific capabilities, scores, or positioning, the capability differences between Pro and Flash, or whether Flash also has omnimodal capabilities.

This means "open-source" is currently more of a publisher's positioning description than a verifiable degree of openness. The license type and whether weights are truly open for download remain unknown.

Comments under the release post show a restrained wait-and-see attitude. A Chinese-speaking user commented that Xiaomi's open-source pace this time is a bit fast, and to wait for real test performance first; another user questioned that if it's just "benchmark gaming" (benchmaxxded), then no one will actually use it.

These two comments themselves are confirmed existing community feedback, but they point to the same gap: the model's actual user experience and community real-test results have not yet appeared.

In the absence of third-party reproduction, waiting and seeing is not conservative, but a reasonable response to the state of evidence.

The evidence comes only from one X platform relay post (TestingCatalog relaying Xiaomi MiMo's release content), with no first-party technical report or model card. All benchmark and ranking data are self-reported by the publisher or relay.

Apart from the release time, the sources do not provide availability, pricing, or deployment information; whether the Artificial Analysis Intelligence Index score of 46 and 6th place ranking have been officially confirmed by that organization is also not stated in the sources.

The criteria and comparison scope for "top open-weight model" are unclear. These gaps determine that any strong conclusion at present can only remain at the level of source claims.