Xiaomi shipped a new model family this week, and the headline spec isn't the parameter count — it's that MiMo-V2.6-Pro and MiMo-V2.6-Flash are both built to handle text, images, video, and audio inside one set of weights, not as separate models stitched together behind a router. Xiaomi is calling that "natively omnimodal," and both models ship under the MIT license with weights open on Hugging Face, meaning anyone can download, fine-tune, and self-host them commercially with no royalty or approval step.
That combination — frontier-class open weights plus four input modalities plus a genuinely permissive license — is rare enough that it's worth a careful look, and big enough of a claim that it's worth checking before repeating. This piece verifies what's actually confirmed this week, flags what's still only Xiaomi's own word, and lays out how Pro, Flash, and a third "UltraSpeed" variant actually compare.
Quick summary: Xiaomi released MiMo-V2.6-Pro, MiMo-V2.6-Flash, and a MiMo-V2.6-Pro-UltraSpeed variant together on September 22, 2026, all MIT-licensed and open-weight on Hugging Face. Pro: a reported 1.02 trillion total parameters, 42 billion active (sparse Mixture-of-Experts). Flash: a reported 309 billion total, 15 billion active. Both are "natively omnimodal" (text, image, video, audio in) with a 1,048,576-token (~1M) context window. Reported API pricing: Pro at $0.435 input / $0.87 output per million tokens, Flash at $0.14 / $0.28. UltraSpeed reportedly runs up to 20x faster than Pro on Xiaomi's own figure (one third-party listing says closer to 10x) at roughly 10x Pro's price. MiMo-V2.6-Pro reportedly scores 46 on the independent Artificial Analysis Intelligence Index, the top mark among open-weight models — the one figure here with real third-party verification behind it. Nearly everything else, including head-to-head claims against Claude Opus 5, is Xiaomi's own launch-week reporting and hasn't been independently reproduced yet.
What "natively omnimodal" actually means
Most "multimodal" chatbots today are a text-trained language model with a vision encoder — and sometimes an audio module — bolted on afterward, trained separately and stitched in later so the model can describe an image or transcribe audio without ever really reasoning across modalities the way it reasons across text. Xiaomi's pitch for MiMo-V2.6 is that omnimodal capability was trained in from the start, across the same reinforcement-learning pipeline, rather than added as a late-stage adapter.
The specs back that framing up in at least one concrete way: reporting on the release describes MiMo-V2.6-Pro as pairing its sparse Mixture-of-Experts language backbone with a 681-million-parameter vision encoder that itself includes a 308-million-parameter audio tokenizer and a 127-million-parameter audio patch encoder — dedicated components for turning sound into tokens the same transformer reasons over, rather than a separate speech-to-text step feeding the model a transcript (Source: SiliconANGLE). Both Pro and Flash accept text, image, video, and audio as input, and both carry a 1,048,576-token (roughly 1 million token) context window — enough to hold long video transcripts, full codebases, or extended multi-turn agent traces in a single call (Sources: llm-stats.com — Pro, llm-stats.com — Flash).
It's worth being precise about what's still unverified: "natively omnimodal" is Xiaomi's own characterization, repeated by outlets covering the launch, not a claim any outside lab has independently tested — for instance, by probing whether the model's video and audio understanding degrades gracefully the way native training would predict, versus how a well-integrated bolted-on adapter would behave. A week after launch, nobody outside Xiaomi has published that kind of teardown yet.
Pro, Flash, and UltraSpeed: the numbers
Xiaomi released three variants together. MiMo-V2.6-Pro is the flagship: a sparse Mixture-of-Experts model with a reported 1.02 trillion total parameters and 42 billion active per token. MiMo-V2.6-Flash is the smaller, cheaper sibling at a reported 309 billion total parameters and 15 billion active. A third variant, MiMo-V2.6-Pro-UltraSpeed, runs the same underlying weights as Pro but through an inference setup Xiaomi says delivers up to 20 times the output speed at similar quality — aimed at latency-sensitive use cases like live trading signals or real-time voice agents, at a materially higher per-token price. Xiaomi also quietly released a smaller, roughly 9-billion-parameter distilled model alongside the three (Sources: llm-stats.com — Pro, OpenRouter — Flash, mimo.mi.com — Pro-UltraSpeed).
| Model | Total / Active Params | Context Window | Price (in / out per 1M tok) | Release Date |
|---|---|---|---|---|
| MiMo-V2.6-Pro | 1.02T / 42B | 1,048,576 tok (~1M) | $0.435 / $0.87 | Sept 22, 2026 |
| MiMo-V2.6-Flash | 309B / 15B | 1,048,576 tok (~1M) | $0.14 / $0.28 | Sept 22, 2026 |
| MiMo-V2.6-Pro-UltraSpeed | 1.02T / 42B (Pro weights) | 1,048,576 tok (~1M) | ~$4.35 / $8.70 (reported) | Sept 22, 2026 |
One correction worth flagging: an earlier brief we worked from suggested Flash shipped a day before Pro, on September 21. Xiaomi's own model-update log lists all three variants — Pro, Flash, and Pro-UltraSpeed — under a single release date, September 22, 2026, with no staggered rollout, so that's the figure this piece uses (Source: mimo.mi.com model updates).
The UltraSpeed speed claim isn't perfectly consistent across sources either: Xiaomi's own product page advertises "up to 20x" faster output than standard Pro at the same quality, while at least one third-party model catalog lists the gain closer to 10x. Both agree UltraSpeed is meaningfully faster than Pro at a premium price — reported around $4.35 input / $8.70 output per million tokens, roughly 10 times Pro's own rate — but the exact speed multiplier depends on whose number you use, so treat "20x" as Xiaomi's marketing figure rather than an independently measured one (Source: OpenRouter — Pro-UltraSpeed).
On Xiaomi's own vendor benchmarks, Flash reportedly stays within about 4 points of Pro across most published tasks — and even edges ahead on one, a coding-security benchmark called CyberGym (95.1 vs. 94.0). If that holds up under independent testing, Flash at roughly a third of Pro's per-token price is the more interesting model for most production workloads, not the flagship.
How good is it, really? The benchmark question
MiMo-V2.6-Pro reportedly scores 46 on the Artificial Analysis Intelligence Index — a composite benchmark run by Artificial Analysis, an independent evaluator, not Xiaomi. We checked Artificial Analysis's own model page directly: it lists MiMo-V2.6-Pro in the mid-40s, ranked first among the open-weight models it tracks and well above the roughly 18-point median for open-weight models of comparable size (Source: Artificial Analysis). Multiple outlets covering the launch describe that score as putting MiMo-V2.6-Pro in the same range as xAI's Grok 4.7 and ahead of other recent open-weight releases like Kimi K3 and GLM-5.3 on that specific index (Sources: officechai, VentureBeat) — though index scores get recalculated as the benchmark suite evolves, so older competitor scores aren't a perfectly clean apples-to-apples baseline months later.
That's the one number in this piece with real independent verification behind it, because Artificial Analysis runs its own evaluation rather than republishing Xiaomi's self-reported figures. Most of the rest of the benchmark story right now is Xiaomi grading its own homework. Coverage of the release cites Xiaomi's own head-to-head comparisons against Claude Opus 5 on benchmarks like AutomationBench (reportedly 53.1 for MiMo-V2.6-Pro vs. 50.3 for Opus 5) and Terminal-Bench 2.1 (reportedly 89.9 vs. 89.1) — narrow, single-digit leads, from a single source, on a model that launched days ago (Source: SiliconANGLE). Narrow leads reported by the model's own maker, on a launch this fresh, are exactly the kind of number that deserves a "reportedly" attached until a neutral third party reproduces them.
One genuinely notable and independently checkable piece of the release: Xiaomi says it's publishing the reinforcement-learning environments it used to train these models — reportedly around 7,000 verifiable coding, cybersecurity, knowledge, and visual tasks plus roughly 1,000 music-generation tasks, each with an automated grader — along with the training code, specifically so outside researchers can attempt to reproduce the results. Xiaomi also disclosed rough compute-spend figures for the RL runs themselves: reportedly about $2.62 million for Pro's six-day run and around $0.85 million for Flash's. If those environments hold up to outside scrutiny, that's a more substantive transparency move than most frontier-adjacent labs make at launch, open-weight or not (Source: mixed-news).
Not the model LumiChats covered before
If "MiMo" sounds familiar, that's because Xiaomi's open-weight line has moved fast this year. We've previously covered the original MiMo V2 Pro against Claude Opus 4.7 on coding and agentic benchmarks, and MiMo-V2.5-Pro's aggressive pricing against Claude Sonnet 4.6. Both of those posts are about earlier generations — text-and-code-focused, agentic-RL-trained models without native image, video, or audio input. MiMo-V2.6 is a different generation on a different axis: the headline change isn't primarily a benchmark jump over V2.5, it's that omnimodality — plus roughly double V2.5-Pro's reported parameter count — arrived in the same release. We're not re-running those older cost comparisons here; if you're weighing Claude against an earlier MiMo generation on price, those posts still apply. This one is about what's new in V2.6 specifically.
Who this actually matters for
- Teams that need to self-host for data residency or compliance reasons and want a model that doesn't require separate vision/audio pipelines bolted on afterward.
- Developers running agentic pipelines who were already comparing MiMo V2.5's price-to-performance against Claude or GPT API rates and want to know if V2.6 changes that math — Flash, at a reported $0.14 / $0.28 per million tokens, is the one to benchmark first.
- Researchers interested in the released RL environments themselves, independent of whether they ever run the resulting model.
- Anyone weighing a genuinely multimodal open-weight model against closed frontier options like Claude Opus 5, Gemini, or GPT-6 for a workload that mixes text, images, and audio in the same pipeline.
It's less compelling for anyone who just wants a polished consumer chat app — Xiaomi's own AI Studio, MiMo Desktop, and MiMo Code surfaces exist, but the mainstream chat-app ecosystem around MiMo is nowhere near as built out as Claude, ChatGPT, or Gemini's, and self-hosting a trillion-parameter sparse MoE model is not a casual undertaking.
01Is MiMo-V2.6-Pro actually free to use commercially?
Yes — both Pro and Flash ship under the MIT license with weights on Hugging Face, so you can self-host, fine-tune, and deploy commercially without paying Xiaomi or requesting permission. Using Xiaomi's own hosted API instead costs a reported $0.435 input / $0.87 output per million tokens for Pro, or $0.14 / $0.28 for Flash.
02What does "natively omnimodal" mean, and is it actually different from other multimodal AI?
It means the model's vision, audio, and video understanding were reportedly trained as part of the same pipeline as its text training, using dedicated encoder components — Xiaomi cites a 681M-parameter vision encoder with built-in audio tokenizer components for Pro — rather than a text model with a vision or speech module attached afterward. That's Xiaomi's own framing; no independent lab has yet published a teardown confirming it behaves meaningfully differently from a well-integrated bolted-on approach.
03How does MiMo-V2.6-Pro compare to Grok 4.7 or Claude Opus 5?
On the Artificial Analysis Intelligence Index — an independent benchmark, not Xiaomi's own — MiMo-V2.6-Pro reportedly scores in the mid-40s, in the same range as Grok 4.7 and ahead of other recent open-weight models, making it the top-ranked open-weight model on that index as of launch week. Xiaomi has also published its own head-to-head numbers against Claude Opus 5 on agentic benchmarks showing narrow leads, but those are vendor-reported and haven't been independently reproduced yet.
04Is this the same MiMo model LumiChats compared to Claude before?
No. Earlier LumiChats posts covered MiMo V2 Pro and MiMo-V2.5-Pro, both text-and-code-focused agentic models without native image, video, or audio support. MiMo-V2.6 is a newer generation that adds omnimodal input and roughly doubles Pro's parameter count over V2.5.
05Should I trust the benchmark numbers this early?
Treat them cautiously. The Artificial Analysis Intelligence Index score comes from an independent evaluator and is the most solid number here. Most of the rest — including comparisons against Claude Opus 5 — comes from Xiaomi's own launch materials, published within the same week as the release, and hasn't yet been checked by outside researchers.
If you want to see how MiMo-V2.6-Pro or Flash actually respond next to Claude, Gemini, or GPT-6 on a prompt you care about, LumiChats lets you chat with a range of leading AI models side by side in one place — useful for judging a same-week launch like this one against your own workload, rather than against vendor-reported benchmark tables alone.
