Xiaomi's MiMo V2.6 Pro becomes top open-weight AI model
What's the story
Xiaomi, a leading Chinese electric vehicle and consumer electronics maker, has launched the latest version of its MiMo language models. The new model, MiMo-V2.6-Pro, has been ranked as the best open-weight model in the world by Artificial Analysis's Intelligence Index. It scored 46 on their scale, beating proprietary models like xAI's Grok 4.6 (44), Google's Gemini 3.8 Flash (41), and DeepSeek V4.1 Flash (39).
Accessibility
MiMo-V2.6-Pro available for free download on Hugging Face
What makes MiMo-V2.6-Pro even more impressive is its open-weight, MIT-licensed status and affordability via Xiaomi's API.
This means independent developers and businesses can download it for free from Hugging Face, customize it for their specific needs, run it on their own or rented hardware, and use it in production workflows without paying anything to Xiaomi.
New release
'MiMo-V2.6-Flash' also launched
Along with MiMo-V2.6-Pro, Xiaomi has also launched MiMo-V2.6-Flash.
This is a smaller and cheaper model ($0.14/$0.28 USD per million input/output tokens) that offers the same 1-million-token context window and native multimodal capabilities as its more expensive counterpart.
It is aimed at high-volume production workloads and is the second cheapest major frontier model available globally through its application programming interface (API).
Speed boost
'MiMo-V2.6-Pro-UltraSpeed' variant also launching
Xiaomi is also launching MiMo-V2.6-Pro-UltraSpeed, a variant that can generate at up to 20 times the normal Pro output speed.
This move is part of Xiaomi's larger strategy to build an entire open agent stack with foundation models, coding agents, harnesses, reinforcement-learning environments, and the training infrastructure behind them.
Growth
Xiaomi's push into frontier models over the past year
Xiaomi's foray into frontier models has picked up pace over the last year. The company started publicly expanding the MiMo family in 2025 and has been pushing deeper into agentic AI during much of 2026.
Its April release of MiMo-V2.5 and V2.5-Pro established many architectural and economic concepts now seen in V2.6: sparse mixture-of-experts models, million-token context windows, permissive licensing, and low prices for long-running agent workloads.
Training details
Training costs for MiMo models
MiMo-V2.6-Pro is a 1.02-trillion-parameter mixture-of-experts model with 42 billion parameters active during inference.
Xiaomi says both MiMo-V2.6-Pro and the smaller MiMo-V2.6-Flash underwent 30 large RL steps covering some 750,000 trajectories in less than six days, at reported costs of about $2.62 million for Pro and $850,000 for Flash.