Xiaomi released the MiMo-V2.6 model family on September 22, 2026, with its flagship MiMo V2.6 Pro scoring 46 on the Artificial Analysis Intelligence Index, first among 114 models in its class and the highest score any open-weight model has reached1,2.
The score puts MiMo V2.6 Pro ahead of Moonshot AI's Kimi K3 (44) and Zhipu's GLM-5.3 (45). Kimi K3 and Alibaba Group Holding's Qwen3.8 Max have traded the open-weight lead for most of this year, and Xiaomi says MiMo V2.6 Pro now passes both.
Architecture and pricing
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 accepts text, images, audio, and video and holds a one-million-token context window. Xiaomi added 3D spatial reasoning and direct computer use to this generation.
Pricing is the sharpest edge. Artificial Analysis rates MiMo V2.6 Pro the cheapest model it tracks, at $0.13 per task. Xiaomi charges $0.435 per million input tokens and $0.87 per million output tokens. For comparison, Anthropic cut Claude Opus 5.5 to $4 input and $20 output on the same day.
The lighter MiMo V2.6 Flash RL costs $0.14 per million input tokens and $0.28 per million output tokens, roughly a third of Pro. Flash keeps the same million-token context and multimodal input as Pro. On most of Xiaomi's own agent benchmarks, Flash trails Pro by four points or fewer. On CyberGym, Flash beats Pro, 95.1 to 94.0.
A third variant, Pro-UltraSpeed, generates up to 20 times faster than Pro at the same quality, according to Xiaomi, but costs $4.35 per million input tokens and $8.70 per million output tokens.
Training economics
Xiaomi says the reinforcement learning phase lasted under six days and generated 750,000 trajectories across 30 training steps, using a fully asynchronous GRPO algorithm distributed over 1,568 prompts and 16 parallel rollouts. The declared training cost was approximately $2.62 million for Pro and $850,000 for Flash. Xiaomi mixed coding, general agents, visual work, and cybersecurity into one training run and stripped build caches and future Git history from the training environments.
On DeepSWE, Xiaomi reports Pro climbing from 58.4 to 72.57 across the run, and Flash from 48.8 to 65.68.
Where it falls short
Closed-weight frontier models still lead on several benchmarks. Artificial Analysis scores Claude Opus 5.5 at 58 versus MiMo V2.6 Pro's 46. Claude Opus 5 beats Pro on DeepSWE and ProgramBench. On Terminal Bench 4.0, GPT-6 Astra scores 59.6 and Claude Opus 5 scores 49.0 against Pro's 34.9. Xiaomi says Pro performs on par with Claude Opus 5 and GPT-5.6 Sol across most agent benchmarks.
Both models are live now on Xiaomi's AI Studio, MiMo Code, MiMo Desktop, Xiaomi's API platform, and OpenRouter. Weights are downloadable from Hugging Face Inc. under an MIT license. Xiaomi also published the full technical report, the end-to-end reinforcement learning framework, more than 7,000 RL task environments with automatic graders, composable mini-harnesses, reward design, hyperparameters, data mixtures, and costs.
ANALYSIS The pricing gap between MiMo V2.6 Pro and Claude Opus 5.5 is roughly an order of magnitude on input tokens and more than 20x on output tokens. For workloads where open-weight hosting and cost matter more than peak benchmark performance, MiMo-V2.6 resets the price-performance floor.