MiMo-V2.6-Pro

Forkast provides an independent technical write-up of the MiMo-V2.6-Pro launch on 2026-09-22, confirming the model uses a frozen-router Mixture-of-Experts architecture with 1.02 trillion total parameters and 42 billion active parameters. The piece also reports hybrid attention, a 5-layer MTP speculative decoder, and fu The article lists vendor-reported benchmarks for MiMo-V2.6-Pro, including 71.9% on DeepSWE v1.1, 89.9% on Terminal Bench 2.1, 94.0% on CyberGym, and 53.1% on AutomationBench v1.0.6; readers are reminded these are vendor-reported. The article also identifies Luo Fuli, formerly of DeepSeek, as the team lead. Coverage inc

The BenchLM aggregator page for MiMo-V2.6-Pro explicitly names the exact model version and tracks published metrics as of the release date. It reports a 1-million-token context window, $0.43 input and $0.87 output pricing, and a cached-input price of $0.004 with a blended figure of $0.65. The page notes 14 published be For evidence coverage, the page shows 9 of 9 verified agentic benchmark rows and 3 of 3 verified coding benchmark rows, while reasoning, multimodal, knowledge, multilingual, instruction-following, and math categories are listed as not yet measured. Independent runtime speed and time-to-first-token are also not measured

Xiaomi officially released and open-sourced the MiMo-V2.6 series, comprising two native fully multimodal models named Pro and Flash, as stated on its own docs/news page. The release centers on scaling reinforcement learning on complex verifiable tasks; the team describes a 6-day "Live RL" training run built on top of r The V2.6-Pro is positioned as the flagship of the family, built for professional workflows and produced via large-scale multi-task RL. The same page notes that MiMo-V2.6-Pro leads domestic open-source models in allocated RL compute and sets a new cost-performance record for the same intelligence level. This is the prim