The new MiMo-V2.6 series from Xiaomi provides open weights for its Pro and Flash models to support agentic AI workloads. The Pro version is a mixture-of-experts architecture with 1.02T total parameters and 42B active parameters, posting a score of 46 on the Artificial Analysis Intelligence Index. This model costs $0.435 per 1M input tokens and $0.87 per 1M output tokens. The release package also includes a technical report, a distilled Qwen 9B model, and an end-to-end reinforcement learning framework with 7,000 task environments.

The MiMo-V2.6 series uses Grouped Query Attention and Sliding Window Attention with a 128 token window to optimize efficiency. Developers increased intelligence by replacing simple correctness verifiers with agentic graders that analyze execution traces and by using RL batches of 25,088 trajectories per update. The model supports a 1M token context and integrates with the vLLM and SGLang inference engines.