MiMo-V2.6-Pro-UltraSpeed: Xiaomi Shipped the Family and Put a 10x Price on Speed

MiMo-V2.6-Pro-UltraSpeed is the same 1.02-trillion-parameter checkpoint as MiMo-V2.6-Pro, served faster, and it costs about ten times as much to call: $4.35 per million input tokens and $8.70 per million output tokens against the standard Pro lane at $0.44 and $0.87. That ratio is the whole story of Xiaomi's September 22 release, and it is a stranger story than the launch coverage suggests — because the thing Xiaomi shipped that nobody had a week ago is not a model. It is a price list for latency.

The models are real now. Xiaomi's launch page is dated September 22, the Hugging Face repositories for the MiMo-V2.6 reinforcement-learning checkpoints went up the day before, and the family — MiMo-V2.6-Pro, MiMo-V2.6-Flash, and the UltraSpeed serving tier — is listed across commercial catalogues with rates attached. Three weeks ago none of that existed. Our own write-up on September 21 said, accurately, that MiMo-V2.6-Pro had no release date, no model card, no API identifier, no price and no downloadable weights. Every one of those five gaps closed inside about 48 hours. So the interesting question is no longer whether Xiaomi shipped. It is what you are actually buying when you pay ten times the standard rate for the same weights.

What a "speed tier" is, and what it is not

UltraSpeed is not a smaller model, a distilled model, or a differently trained model. Xiaomi describes it as the fast edition of the flagship, built from the same 1T MiMo-V2.6-Pro checkpoint, and the pricing catalogues carry the same 1M-token context window and the same native multimodality. There is no new capability in the tier. What changes is how the checkpoint is served: the batching, the speculative decoding, the hardware allocation, the concurrency limits — the engineering between the weights and your HTTP request.

That distinction matters more than it sounds, because it determines what you can and cannot conclude from a speed tier's existence. A serving-optimised variant tells you Xiaomi believes there is a market of buyers for whom token latency is worth more than token cost. It does not tell you the model got better, and it does not tell you the weights changed. If you benchmark UltraSpeed and the standard Pro lane on the same prompts, you should expect the same answers arriving at different rates. If you get different answers, that is a finding worth reporting, not a feature.

The 20x claim and the 10x claim

Here is where the sourcing gets interesting, and where most of the launch coverage has quietly picked a number without saying which one it picked. Xiaomi's own material claims UltraSpeed reaches up to 20 times the output speed of the standard Pro service at the same quality. Third-party catalogue listings, describing the same model on the same day, say it delivers roughly 10 times the output speed. Both numbers are in circulation. They are not the same number, and nobody has published a measurement that reconciles them.

We are not going to pretend to resolve it. The honest reading is that "up to 20x" is a vendor ceiling under conditions Xiaomi has not specified — a best case, probably at favourable batch sizes — while "roughly 10x" is what a catalogue listing was willing to assert as a typical figure. Until someone publishes per-request latency distributions at a stated concurrency level, treat the real multiplier as somewhere in a wide band and test it on your own traffic. What is not in dispute is the direction: the tier is materially faster, and it is priced as if it were.

The arithmetic of paying for speed

The price ratio and the speed ratio do not match, and that is the part worth doing arithmetic on. If UltraSpeed is 10x faster and 10x more expensive per token, then for a fixed amount of work — a fixed number of tokens generated — you pay 10x more and finish in a tenth of the wall-clock time. Cost per token is up 10x. Cost per unit of time saved is the thing you are actually buying, and it has no benchmark.

• Standard MiMo-V2.6-Pro — $0.44 in / $0.87 out per 1M tokens • MiMo-V2.6-Pro-UltraSpeed — $4.35 in / $8.70 out per 1M tokens • Ratio — roughly 9.9x on input, exactly 10x on output • Claimed speed gain — up to 20x (Xiaomi), roughly 10x (third-party listings) • Capability difference — none claimed; same checkpoint, same 1M context, same multimodal inputs • Licence — MIT on the released MiMo-V2.6 weights, per Xiaomi's own model cards

Whether that trade is worth taking depends entirely on what the latency is attached to, and the answer flips depending on the workload. For a batch job — overnight document processing, offline evaluation, dataset generation — latency is nearly worthless and the standard Pro lane is the obvious call. For an interactive agent loop where each turn gates the next, latency compounds: a 10x faster model in a 30-step agent run is not 10x faster end to end, but it is the difference between a session a user waits through and a session a user abandons. Xiaomi is plainly betting that the second kind of buyer exists and will pay.

What Xiaomi actually put on the record

The released family is a sparse mixture-of-experts design. Xiaomi's own model card for the MiMo-V2.6 RL checkpoints describes the series as 1.02T total parameters with 42B activated per token, a 1M-token context window, native multimodality across text, image, video and audio, and an MIT licence tag on the public repositories. The release bundle includes a technical report and deployment notes, and Xiaomi says it is open-sourcing the reinforcement-learning training environment and code so the post-training can be examined and reproduced. That is a wider disclosure than an API launch, and it is a genuine step beyond what was visible when the company was only streaming a training dashboard.

The benchmark numbers remain vendor-reported and unreproduced. On DeepSWE v1.1 — Xiaomi's own harness, mini-swe-agent, avg@3 — the company reports MiMo-V2.6-Flash moving from 48.8 to 65.68 and MiMo-V2.6-Pro from 58.4 to 72.57 across the reinforcement-learning run. The post-training itself is documented as 30 steps and roughly 750,000 trajectories per model, completed in under six days, at a combined published spend of about $3,474,715 across the Pro and Flash runs. Those figures come from Xiaomi's own telemetry and its own evaluation harness. No independent party has reproduced the scores, and the distinction between "Xiaomi's harness says 72.57" and "72.57 is the score" is the difference between a claim and a fact. A third-party tracker additionally lists the Pro model at 46.32 on an Artificial Analysis intelligence index; we have not read that figure off Artificial Analysis directly, so treat it as a pointer rather than a number.

One caution about the MIT tag, because it is the kind of detail that gets flattened in launch coverage. The licence applies to the weights Xiaomi published. The UltraSpeed tier is a hosted service, and a hosted service is governed by terms of service, not by a weight licence. Having the right to run the checkpoint on your own hardware and having the right to resell someone's accelerated serving of it are different rights, and only the first one comes with the MIT tag.

Getting to it, and what we do and do not route

MiMo-V2.6-Pro-UltraSpeed is available through the vendor's own API and several third-party platforms. It is not on OrcaRouter — no Xiaomi model is, and we would rather say that plainly than let a model page imply otherwise. If you are calling it, you are calling it through Xiaomi or through one of the commercial catalogues carrying it.

Where a routed catalogue earns its place with a model like this is in the trying, not the hosting. A brand-new premium tier with a 10x price tag and a speed claim nobody has independently measured is exactly the kind of thing you want to A/B against your incumbent before you commit a production path to it — and doing that by standing up a second contract, a second key and a second billing relationship is the friction that stops teams from ever running the test. OrcaRouter puts 200+ models behind one OpenAI-compatible endpoint at 0% markup, passing provider list price straight through, with automatic failover across providers. That means a vendor price change is live on our side the same day rather than at the next billing cycle, and it means the comparison you want to run — UltraSpeed against the model it would replace, on your own prompts — is a config change rather than a procurement conversation. If Xiaomi cuts the UltraSpeed rate next month, as vendors usually do with a new premium tier, the number on our side moves with it.

What to watch

Three things would settle the open questions in this release. An independent latency measurement at a stated concurrency would tell you whether the real multiplier is nearer 10x or 20x, and whether it holds under load rather than in a single-stream demo. An independent reproduction of the DeepSWE figures would tell you whether 72.57 is a model property or a harness property. And the first price movement on the UltraSpeed tier would tell you whether Xiaomi priced it as a permanent premium or as a launch number — the standard Pro lane at $0.44/$0.87 is the anchor, and a tier that started at 10x the anchor rarely stays there.

Until then, the defensible position is narrow and clear. The family shipped, the weights are MIT and downloadable, the technical report exists, and the standard Pro lane is a genuinely cheap way to call a 1T-class model. UltraSpeed is a bet on latency, priced at a premium that only makes sense for interactive and agentic loops, and carrying a speed claim with a 2x spread between the vendor's ceiling and everyone else's description. Buy it if the wall-clock time is the product. Do not buy it expecting a better model.