Updated on September 28, 2026
MiMo-V2.6 is Xiaomi's open AI model series, three MIT-licensed variants that accept text, images, video and audio across a 1-million-token context window. Released on September 22, 2026, the family spans the trillion-parameter Pro down to the budget-minded Flash, with API rates from $0.14 per million input tokens. Weights, training code and over 7,000 RL environments ship alongside, a rare package for developers building agents or hosting models themselves.
The series includes three variants, all presented on September 22, 2026, with text, image, video and audio input and a 1-million-token window. Xiaomi calls them natively omnimodal, with spoken output handled by its MiMo-V2-TTS module.
The release goes well beyond a weights drop. The model family launched in April 2025 this time ships the technical report, the training code and more than 7,000 reinforcement learning environments, everything under an MIT license. A 9-billion-parameter distill built on Qwen3.5-9B was also announced for modest machines.
MiMo-V2.6-Pro scored 46 on the Artificial Analysis Intelligence Index at launch, the highest mark an open-weight model has ever reached, tied with Grok 4.7 released the same day. Xiaomi's own figure is 46.32, and the company claims a lead over Kimi K3 and Qwen3.8 Max.
Closed frontier models keep some distance, with Claude Opus 5.5 around 58 and GPT-6 Sol near 48 according to press coverage. On the DeepSWE v1.1 software benchmark, Xiaomi reports 72.57 for Pro, up from 58.4, and 65.68 for Flash. Company numbers, so treat them as such (the final training run reportedly cost $2.62 million, pocket change by frontier standards).
The API bills per million tokens, at the same rates as the V2.5 generation. You grab a key on the MiMo API Platform or go through OpenRouter, and existing scripts run almost untouched since the call format matches the usual clients.
Cache hits drop Flash input to $0.0028 per million. Rates shift as Xiaomi updates the range, so the official platform holds the current grid. To place the series among current large language models, the short version is that the weights sit on Hugging Face and that AI Studio, MiMo Code and MiMo Desktop already carry V2.6.
| Model | Input / M tokens | Output / M tokens | Positioning |
|---|---|---|---|
| MiMo-V2.6-Flash | $0.14 | $0.28 | Balance of intelligence and cost |
| MiMo-V2.6-Pro | $0.435 | $0.87 | The most capable of the series |
| MiMo-V2.6-Pro-UltraSpeed | $4.35 | $8.70 | Up to 20x faster output |
Xiaomi MiMo is the company's family of language models, started in April 2025 with MiMo-7B and led by Luo Fuli, formerly of DeepSeek. It also powers Xiaomi's phone, car and home ecosystem. MiMo-V2.6, released in September 2026, is the current generation, with Pro, Flash and Pro-UltraSpeed variants.
Pro packs 1.02 trillion parameters (42B active) against 309 billion (15B active) for Flash, which costs roughly a third as much per token. Xiaomi aims Pro at long, complex projects and Flash at frequent calls and heavy workloads. On the company's own agent tests, Flash trails Pro by only a few points.
Yes, both Pro and Flash are listed on OpenRouter with the full context window, billed at Xiaomi's standard rates. The well-known ':free' route concerns the older MiMo-V2-Flash, not V2.6. Downloading the weights from Hugging Face costs nothing, though serving a model this size takes serious hardware.
Coding agents lead the list, with Xiaomi reporting 72.57 on the DeepSWE v1.1 software benchmark for Pro. Training also covered computer use, cybersecurity, office work and visual tasks, and the million-token context plus low rates suit long tool-calling loops. MiMo Code and the published mini-harnesses provide ready starting points.
Verdict: Forty-six points on an independent index at a fraction of frontier prices, that is the pitch in one line. MiMo-V2.6 suits developers running token-hungry agents, teams hosting their own weights, and researchers keen to dissect a fully published RL stack.
