Xiaomi open-sources MiMo-V2.6 Pro and Flash

Xiaomi’s MiMo team released MiMo-V2.6-Pro and MiMo-V2.6-Flash as open omnimodal models — text, image, video, and audio — with Hugging Face weights, a technical report, and disclosed reinforcement-learning tooling, while Xiaomi’s platform changelog dated the series for the same news window.

SOFTWARE desk — another large Chinese open omnimodal stack with downloadable weights and RL machinery means developers can test claims instead of trusting a press chart alone.

What Pro is, on the card XiaomiMiMo/MiMo-V2.6-Pro-RL. It is a sparse mixture-of-experts model, or MoE: a large set of specialist blocks, with only some turned on for each piece of input. Xiaomi says it has about 1.02 trillion parameters in total and 42 billion activated. A parameter is one learned number. 1.02 trillion is the whole warehouse. 42 billion is the slice that runs. The card lists a context length of 1 million tokens. A token is a small chunk of text, roughly a word piece. Context length is how much the model can hold at once — a long code folder, a long tool trace, or several sessions. The card says the inputs are text, image, video, and audio in one model. Omnimodal, here, means those four kinds of input, not text alone. The card’s license line is MIT. MIT is a permissive open-source license: others may use and build on the weights under that license’s conditions. File the size, the context, the four inputs, and MIT as the Pro card’s.

Flash is the smaller sibling, on the card XiaomiMiMo/MiMo-V2.6-Flash-RL. Xiaomi says 309 billion parameters in total and 15 billion activated, also a sparse MoE, also a 1-million-token context, also text, image, video, and audio. The license line is MIT, the same as Pro. Xiaomi’s updates page calls Flash a full-modality, high-intelligence, low-cost reasoning model, and “the best balance for high-frequency calls and large-scale tasks.” “Best balance” and “low-cost” are Xiaomi’s wording. File the 309 billion / 15 billion split as the Flash card’s. This desk did not compare a bill.

Pro-UltraSpeed is a hosted tier on that same updates page, not a third weight file this desk found. Xiaomi says it is flagship V2.6-Pro performance, up to 20 times faster, built for real-time work. Up to 20 times faster is Xiaomi’s claim. RuntimeWire reads the tier as a higher-priced hosted option at the same model quality. File the 20-times line as Xiaomi’s, and the higher-price reading as RuntimeWire’s. This desk did not time a reply.

What shipped with the weights. Both the Pro and Flash repositories include a technical-report PDF, named MiMo_V2_6_technical_report.pdf, plus the model card. The card’s title frames the series as scaling reinforcement learning toward self-improvement. Reinforcement learning, or RL, means the model is scored on finished tasks and steered toward better scores. The card says one mixed RL run covers coding, general agents, visual work, and cybersecurity, instead of a separate run for each. It names the batch: 1,568 prompts times 16 tries per step. A prompt is the task. A try, which the card calls a rollout, is one attempt. The card also describes graders that rank successful tries against each other, because a simple pass-or-fail cannot tell a short good answer from a long one that happened to pass. RuntimeWire says Xiaomi is also opening the training environments and the RL code needed to inspect and reproduce the work, and that each of Pro and Flash finished 30 RL steps across roughly 750,000 trajectories in fewer than six days. A trajectory is one full attempt, start to finish. File the mixed run, the batch size, and the grader idea as the card’s. File the environments line, the 30 steps, the 750,000 trajectories, and the six-day line as RuntimeWire’s account of Xiaomi’s launch page. This desk did not open that launch page, did not clone a training repository, and did not read the PDF as a second set of numbers.

A smaller checkpoint is also on Hugging Face: MiMo-V2.6-Distill-Qwen-9B, created the same day. Xiaomi says it is a 9-billion-parameter model made by supervised fine-tuning of Qwen3.5-9B on data the MiMo models generated. Supervised fine-tuning means training on example answers, not the big RL run. Qwen is Alibaba’s open model family. The card says this checkpoint is a starting point for open research in agentic reinforcement learning — RL for software that takes a series of steps, not only a chat reply. It covers coding, general agent tasks, visual coding, and cybersecurity. It is not the 1.02-trillion flagship. File that as the distill card’s.

The score claim stays labeled. RuntimeWire says Xiaomi says Pro scored 46.32 on version 4.3 of the Artificial Analysis Intelligence Index, and that this places it ahead of other open models in that comparison. Artificial Analysis is an outside scoreboard. The claim diagram on this filing rounds that figure to 46. It is Xiaomi’s number, via RuntimeWire. It is not a score this desk measured, and it is not a Bad Signal ranking. The Pro card’s own table is mixed, and it is Xiaomi’s measurement. On DeepSWE v1.1, a software-engineering test, Xiaomi lists Pro at 71.9 and Flash at 67.9, next to Claude Opus 5 at 74.0 and GPT-5.6 Sol at 73.0. On Terminal Bench 4.0, Xiaomi lists Pro at 34.9, behind Claude Opus 5 at 49.0 and GPT-5.6 Sol at 39.9. On AutomationBench, Xiaomi lists Pro at 53.1, ahead of those two closed models on that one row. File those cells as Xiaomi’s. A reader should not take the table as Pro beating Claude or GPT overall.

Hosted access, still the cards. Xiaomi says the models are also available in AI Studio, MiMo Code, Xiaomi MiMo Desktop, the Xiaomi MiMo Open Platform API, and OpenRouter. An API is a hosted call over the internet. Open weights are the downloadable files. Hosted access is a service you call. The open weights are the fact this desk is filing. RuntimeWire also says the series keeps the standard API prices used for MiMo-V2.5. This desk did not open a price table, so that price line stays RuntimeWire’s. There is no dollar figure in this filing.

Leave the Artificial Analysis line as a company claim. The published RuntimeWire article this desk read does not print a training-cost dollar figure, so this filing does not invent one. Leave the 20-times speed line as Xiaomi’s claim, not a time this desk measured. This desk did not download a trillion-parameter checkpoint.

Plain English for the rest of the card: MoE / mixture-of-experts = a model with many specialist blocks, only some active at once. parameter = one learned number inside the model. token = a small chunk of text. context length = how much the model can hold at once. omnimodal = text, image, video, and audio in one model. MIT = the open-source license on the Pro and Flash cards. reinforcement learning / RL = training that scores finished tasks and steers the model toward better scores. rollout = one attempt at a task. API = a hosted call over the internet. open weights = the downloadable model files. hosted tier = a service you call, rather than a file you download. This filing is the open-weight release. It is not a score this desk reran.

PRIMARY here: Xiaomi MiMo’s platform model-updates page and the Hugging Face cards for XiaomiMiMo/MiMo-V2.6-Pro-RL and XiaomiMiMo/MiMo-V2.6-Flash-RL — Tier A PRIMARY, the original record — plus the distill card XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B. RuntimeWire’s 21 Sep 2026 article is the same-day news account, not a second company newsroom. The 1.02-trillion / 42-billion split, the 1-million-token context, the text-image-video-audio inputs, the MIT license, the mixed RL run, the 1,568-by-16 batch, and the grader description are the Pro card’s. The 309-billion / 15-billion split and the matching inputs, context, and MIT line are the Flash card’s. The Pro, Flash, and Pro-UltraSpeed lines, including up to 20 times faster, are the updates page’s, under a 22 Sep 2026 heading. The 9-billion distill of Qwen3.5-9B is the distill card’s. The repository created-at stamps, 21 Sep 2026, are the Hugging Face API’s. The 46.32 Artificial Analysis claim, the 30 steps, the roughly 750,000 trajectories, the fewer-than-six-days line, the training-environments line, the higher-priced UltraSpeed reading, and the V2.5 price line are RuntimeWire’s. NOT claimed: an independent rerun, that Pro leads every closed model, a training-cost dollar figure, that this desk timed the 20-times claim, that this desk opened Xiaomi’s launch page or read the technical report as a second score table, a stock tip, or investment advice. Distinct from the already-filed spacexai-grok-4-7, zai-zcode-open-source, and qwen-image-2.1.