MiMo V2.6 Pro: Xiaomi's Open-Weight Omnimodal Model, Pricing and Specs

What MiMo V2.6 Pro is

MiMo V2.6 Pro is the larger of the two models in Xiaomi's MiMo V2.6 series, announced by Xiaomi on September 22, 2026. Xiaomi describes both V2.6 models as natively omnimodal, and the Hugging Face model card lists text, image, video and audio as input types. The weights are published under the MIT licence, and Xiaomi says it is also releasing the technical report, the training environments and the reinforcement-learning code. The model card describes the published checkpoint as the MOPD upgrade of the MiMo-V2.6-Pro-RL checkpoint, a stage that distills several domain-specialised teacher models into the student on-policy.

Architecture

According to the model card, MiMo V2.6 Pro is a sparse mixture-of-experts model with 1.02 trillion total parameters, of which 42 billion are active per token. It has 70 layers: 60 use sliding-window attention with a window of 128 and 10 use global attention. Each MoE layer has 384 routed experts with 8 activated, and there are no shared experts. The hidden size is 6,144 and the context length is 1M tokens. The card gives serving instructions for SGLang and vLLM, with multi-node settings for the full model.

Pricing

Xiaomi's announcement lists MiMo V2.6 Pro at $0.435 per million input tokens on a cache miss and $0.87 per million output tokens, with cache hits at $0.0036 per million. Cache writes are free for a limited time, according to the same page. Xiaomi states the V2.6 series keeps the API pricing of the V2.5 series. A faster serving mode, MiMo-V2.6-Pro-UltraSpeed, is priced separately at $4.35 input and $8.70 output per million tokens; Xiaomi says it runs at up to 20 times the output speed of the standard mode. RMB pricing for China is listed on Xiaomi's pricing page.

Training run

Xiaomi streamed the reinforcement-learning production run live. Its announcement states that in under six days MiMo V2.6 Pro completed 30 RL steps over roughly 750,000 trajectories at a cost of about $2.62 million, with up to 1M context length during training and 3.5 to 3.7 billion tokens per step. Xiaomi reports that on DeepSWE v1.1, which it describes as a held-out long-horizon software engineering benchmark, the Pro model's score rose from 58.4 to 72.57 over the run.

Benchmarks, and who ran them

On LMArena's text leaderboard MiMo V2.6 Pro has no LMArena rating. Xiaomi's own published figures are its claims rather than independent measurement: it reports a score of 46.32 on Artificial Analysis's Intelligence Index v4.3, which it says places the model above Kimi K3 and Qwen3.8 Max among open-source models. Mistral's Mistral Large 4 announcement lists MiMo-V2.6-Pro among the models it compares against on AutomationBench.

Availability

Xiaomi lists MiMo V2.6 Pro as available in AI Studio, MiMo Code and MiMo Desktop, through the MiMo API Platform, and on OpenRouter. MiMo Desktop left early access with this release and ships with the V2.6 models built in. A Token Plan is offered for high-volume use. The open weights are published on Hugging Face under the XiaomiMiMo organisation, alongside a Flash variant, MiMo V2.6 Flash, and community quantisations exist for both.

Official Resources