Overview
MiMo-V2.6-Pro is the native omni-modal model Xiaomi released and open-sourced on September 22, 2026. It is a sparse mixture-of-experts build with 1.02 trillion total parameters and 42 billion activated per token. Text, image, video and audio all enter the same model, the context window reaches 1M tokens, and a five-layer multi-token-prediction head handles speculative decoding on the output side. Weights ship under the MIT license with no extra conditions for commercial use.
Xiaomi calls the core move of this generation Scaling RL. Rather than post-training for one narrow capability, the team mixed coding, general agent work, vision and cybersecurity into a single reinforcement-learning run while scaling up batch size, the number of environments, and the compute spent on graders. Training ran as a single asynchronous GRPO pass, 1,568 prompts with 16 rollouts each per step. Pro completed 30 steps and roughly 750,000 trajectories in under six days, at a company-reported RL cost of about 2.62 million dollars, and the training framework plus more than 7,000 task environments were released alongside the weights.
On third-party numbers, the Artificial Analysis Intelligence Index v4.3.2 scores it 46, the highest of any open-weight model at launch, against 26 for the previous MiMo-V2.5-Pro. In the Code Arena WebDev arena it measured 1,628, roughly tenth overall and third among open-weight entries, up 153 points from the previous generation. Xiaomi also reports DeepSWE v1.1 rising from 58.4 to 72.6, CyberGym at 94.0 and OSWorld desktop control at 82.0. Those are vendor figures and worth waiting for community reruns before treating them as settled.
API pricing carries over from V2.5 without change: 3 yuan input and 6 yuan output per million tokens, with cached hits at 0.025 yuan. Xiaomi puts the equivalent-intelligence price at one twentieth to one sixtieth of overseas models. The MiMo Desktop client and its membership plan launched at the same time.
Key Features
- Native omni-modal input: Text, image, video and audio all enter one model, no separate encoder pipelines to stitch together.
- Trillion-scale sparse MoE: 1.02T total against 42B activated, separating capacity from per-token inference cost.
- 1M-token context: Built for large repositories, long tool traces and multi-turn agent tasks with fewer mid-run summaries.
- Scaling RL training: Coding, agents, vision and security mixed into one reinforcement-learning run instead of narrow post-training.
- Computer use and 3D reasoning: Official demos cover office software, modeling in Blender, and generating playable 3D open worlds from text or reference images.
- Open weights under MIT: Weights and technical report released, together with the Distill-Qwen-9B model and RL task environments.
Use Cases
- Long-horizon coding agents that edit across files, run tests and recover from failures
- Desktop automation that reads graphical interfaces and completes tasks in productivity software
- 3D and creative production from text or reference images into scenes, models and visual assets
- Research and engineering simulation such as materials screening and Lean 4 proofs that need long reasoning chains
Pros
- The highest intelligence-index score in the open-weight tier, with MIT licensing and no commercial strings
- Native omni-modal input with 1M context removes both modality plumbing and constant summarization
- Pricing unchanged from the previous generation, so the capability gain does not show up on the bill
- Training framework and RL task environments are open, making post-training inspectable and reproducible
Pricing
Token-based API pricing carried over from the V2.5 series: 3 yuan per million input tokens, 6 yuan per million output tokens, and 0.025 yuan for cache hits. Weights are open under MIT and can be self-hosted. Check official listings for current rates and quotas.
Summary
MiMo-V2.6-Pro is Xiaomi next-generation native omni-modal open-weight model: a sparse MoE at 1.02T total and 42B active parameters with a 1M-token context, taking text, image, video and audio into one model. It scores 46 on the third-party intelligence index, the top open-weight result at launch, and 1,628 on the Code Arena WebDev arena. Coding, agents, vision and security were trained in one large reinforcement-learning run whose cost Xiaomi published at about 2.62 million dollars. Pricing carries over from the previous generation.
Version History
- Xiaomi releases and open-sources MiMo-V2.6, Pro tops the open-weight intelligence index (2026-09-22): Xiaomi released and open-sourced the MiMo-V2.6 series, comprising two native omni-modal models, Pro and Flash, which the company frames as a step toward recursive self-improvement. Pro is a sparse MoE with 1.02T total and 42B active parameters, a 1M-token context and MIT-licensed weights. It scored 46 on the Artificial Analysis Intelligence Index, the highest open-weight result at launch, and 1,628 on the Code Arena WebDev arena. Training used large-scale asynchronous GRPO with 1,568 prompts and 16 rollouts per step, finished in under six days, and was reported by Xiaomi at about 2.62 million dollars of RL compute, with over 7,000 RL task environments and the training framework released alongside. API pricing is unchanged from the previous generation, and the MiMo Desktop client and its membership plan launched at the same time.