Xiaomi MiMo-V2.6-Pro Tops Open-Weight AI Leaderboard

Xiaomi has unveiled its MiMo-V2.6-Pro model, which has claimed the top spot on the Artificial Analysis Intelligence Index for open-weight models with a score of 46, up from the 26 scored by its predecessor, MiMo-V2.5-Pro. The new model utilizes a sparse mixture-of-experts architecture featuring 1.02 trillion total parameters, though only 42 billion are active during inference. It supports a one-million-token context window and native multimodal inputs, making it highly optimized for agentic coding, research, and long-horizon tool-use workflows.

The model family includes three versions: Pro, Flash, and UltraSpeed. Xiaomi has priced the API at $0.435 per million input tokens, which drops by 99 percent with a cache discount, and $0.87 per million output tokens, averaging about $0.13 per Index task. For high-throughput needs, the UltraSpeed variant runs roughly 10 times faster than the Pro version, achieving speeds above 1,000 tokens per second on commodity GPUs using TileRT and selective MXFP4 four-bit quantization on its expert layers.

To refine the model, Xiaomi streamed its reinforcement learning post-training live, showing a 30-step process utilizing 1,568 prompts and 16 rollouts to generate 25,088 trajectories per step, costing $2.62 million in compute for the Pro version. For practitioners, this release offers a highly competitive option on the intelligence-versus-cost frontier. While the previous MiMo-V2.5-Pro proved its capabilities on benchmarks like ClawEval, GDPVal, and SWE-bench Pro, developers adopting the new V2.6-Pro will need to carefully evaluate its long-context fidelity, agent reliability, and the hardware demands of hosting a trillion-parameter checkpoint.