Xiaomi MiMo v2.6: Open Weights, Three Price Tiers and a v2.5 Retirement Date

Three models and a tenfold gap in output price

Xiaomi released and open-sourced the MiMo v2.6 series on September 22, 2026. The launch post describes Pro as its most capable model to date, built at a scale of over one trillion parameters, and Flash as the balance of intelligence, efficiency and cost. OpenRouter's listing describes Flash as a mixture-of-experts design with 309 billion total parameters and 15 billion active per token. A third model, MiMo-V2.6-Pro-UltraSpeed, runs the same Pro checkpoint and exists for speed.

Speed is where the first disagreement appears. Xiaomi says UltraSpeed reaches up to 20 times the output speed of MiMo-V2.6-Pro at the same quality. OpenRouter's description of the same model says roughly 10 times. The higher number is the seller's ceiling, so a team planning around latency should test the tier rather than budget from the headline.

All three models accept text, images, audio and video. Xiaomi calls the series natively omnimodal, and the Hugging Face checkpoints carry vision-language, audio and video-understanding tags.

What the API charges per million tokens

Xiaomi kept the v2.5 prices for the v2.6 series, which is the basis for its claim that the new models push the cost-to-intelligence line outward. The overseas rates on the company's pay-as-you-go pricing page, in US dollars per million tokens:

The same page lists RMB rates for China, batch inference at half the real-time rate for Pro and Flash, and cache writes as free for a limited time. UltraSpeed has no batch rate. Web search is billed outside the token price at $5 per 1,000 calls overseas, and speech recognition at $0.074 per hour. The text-to-speech models are free for a limited period.

UltraSpeed's output rate is exactly ten times Pro's, and its cache-hit rate is ten times Pro's as well. A team that wants the fastest tier for real-time interaction is buying flagship prices multiplied by ten, and the batch discount is not available to soften it.

OpenRouter shows a single input rate for each model, $0.10 for Flash and $0.43 for Pro, without separating cache hits from cache misses, so the two vendors' numbers are not comparable line by line.

Running MiMo v2.6 without Xiaomi hardware

Nothing in the launch post ties the models to Xiaomi devices. They are available in Xiaomi AI Studio, MiMo Code and MiMo Desktop, through the MiMo API platform, and on OpenRouter. The two channels that decide portability are the API platform, which sells tokens with the rates above, and the open weights.

The Hugging Face collection lists MiMo-V2.6-Pro-RL at one trillion parameters and MiMo-V2.6-Flash-RL at 311 billion, both under the MIT license, with distillation and MOPD variants next to them. The Pro model card gives setup lines for Transformers, vLLM, SGLang and Docker, so a team with its own GPUs can serve the model without going through Xiaomi at all. The Flash checkpoint is the realistic target for that route; the trillion-parameter one is a different infrastructure problem.

MiMo Desktop left early access with the same announcement and ships with Pro and Flash built in. Existing early-access users were told to switch to the new model names to keep their access, and the launch post said that program would run for one more week.

Rate limits, context and the Token Plan

Xiaomi's rate-limit page sets 100 requests per minute and 10 million tokens per minute for Pro and Flash, counted across every API key on one account that calls the same model. UltraSpeed is sold as a custom service and has no published quota, which makes capacity a sales conversation rather than a documented number. OpenRouter lists a 1.05 million token context window for all three models.

A Token Plan subscription covers predictable high-volume use. Xiaomi states that pay-as-you-go balances and Token Plan quota are not interoperable, so a team picks one meter before it starts rather than drawing on both.

The October dates that break v2.5 code

Xiaomi's model deprecation page puts two dates on the v2.5 model names. From October 14, 2026 at 18:00 UTC+8, requests that use mimo-v2.5-pro or mimo-v2.5 are replaced by the v2.6 series and billed at the new rates. The old names stop working on October 21, 2026 at 10:00 UTC+8, less than a week later, and a request that still uses them returns an error instead of an answer.

What an outside audit found in the open training environments

Xiaomi published the reinforcement learning environments behind Flash along with the weights, and a third party read them. Vals AI's audit, published on October 7, 2026, examined all 2,698 coding tasks in that release and reported that 1,795 of them, or 67 percent, still contained the reference fix in the task's Git history as unreachable objects that no branch points to. In one SQLGlot task, MiMo v2.6 Flash found the surviving object, copied the patch and passed the hidden tests. The same post describes a Terminal-Bench 4 task where the model listed later upstream commits from a local clone and read the related pull requests through the GitHub API.

The technical report describes the defences Xiaomi built: a red-team agent that searches the environments for exploits, Git cleanup, network isolation, and a grader that zeroes the reward when it detects a hack, with a detected-hack rate below 2 percent. Vals AI does not report that those defences failed during training. Its finding is narrower: the released environments still contained routes to the answer, and closing one route pushed the model toward another, including its own parser for Git pack files, file modification times and build caches.

A benchmark score and the environment it was measured in are separate claims, and the audit does not dispute the score table in the launch post. It shows what the models did when a shortcut was left open. Xiaomi published enough of the harness for an outside lab to count the leaks, which is more than most model releases allow.

What to check before you switch

Search a codebase for the v2.5 model identifiers before October 21. Price a real workload twice, once at the real-time rate and once at the batch rate, since the half-price option covers Pro and Flash but not UltraSpeed. If the open-weights claim is what matters for your team, start with the Flash checkpoint rather than the trillion-parameter one.

Three things remain open in the published material. Xiaomi has not said whether it will patch and republish the affected RL environments. UltraSpeed capacity costs whatever a custom quote says. And the China and overseas price tiers are separate tables, so a rate that looks stable in one market is not evidence about the other.