Xiaomi has released a new generation of open AI models, MiMo-V2.6. The flagship MiMo-V2.6-Pro has taken the top spot among models with open weights in the Artificial Intelligence Intelligence Index, scoring 46.32 points and surpassing competitors such as GLM-5.3, Kimi K3, DeepSeek V4.1 Flash, and Qwen3.8 Max. In individual agent tests, the model reached or slightly exceeded proprietary systems: 53.1 in AutomationBench compared to 50.3 for Claude Opus 5, 31.6 in Agents' Last Exam (parity with Opus 5), and 89.9 in Terminal Bench 2.1 compared to 89.1 for Opus 5. In aggregate metrics, Pro still lags behind closed-source flagships – Claude Fable 5.1 and GPT-6 Astra.
Pro is built on a mixture of experts architecture: 1.02 trillion parameters, of which 42 billion are activated per token. The more compact Flash has 309 billion parameters and 15 billion active. Both models are natively omnimodal – processing text, images, video, and audio, supporting up to 1 million tokens of context and available with open weights under the MIT license on Hugging Face. The Pro architecture includes a machine vision encoder with 681 million parameters, an audio tokenizer with 308 million, and an audio patch encoder with 127 million, capable of distinguishing speech.
Pricing through Xiaomi API: Pro – $0.435 per million input tokens without cache, $0.87 per million output, and $0.0036 when hitting the cache. Flash – $0.14, $0.28, and $0.0028 respectively. Executing one task in the Intelligence Index costs an average of $0.13, and the generation speed reaches 125 tokens per second. Models are also available through OpenRouter with 1.05 million tokens of context, AI Studio, MiMo Desktop, and MiMo Code. The Pro-UltraSpeed version provides up to 20 times higher speed at the same quality but costs $4.35 and $8.70 per million tokens.
Improvements over MiMo V2.5 Pro (April) in programming, agent tasks, and cybersecurity are attributed by the company to large-scale reinforcement learning: 30 RL steps and approximately 750,000 task runs, some of which simulated long multi-step processes. Xiaomi has published a technical report, RL code, and training environments. Additionally, a research model MiMo-V2.6-Distill-Qwen-9B based on Qwen3.5-9B has been released to study the transfer of agent capabilities to smaller models.
Related articles
