MiMo v2.6

The top open-weights AI model now costs $0.13 per task, making advanced AI capabilities more accessible.

Xiaomi has released MiMo-V2.6-Pro, an open weights, natively omnimodal AI model. The Pro version is described as their most capable model to date, while a "Flash" variant offers a balance of intelligence, efficiency, and cost. An "UltraSpeed" version provides significantly faster output. This release positions Xiaomi, traditionally known as a phone maker, as a notable player in the frontier AI lab space. The development involved extensive training at an estimated cost of $3 million, focusing on large batches, rich multi-task environments, and advanced reward signals through reinforcement learning (RL). Xiaomi also plans to open-source the RL tooling, environments, and training recipes, though the full datasets are not yet public. The MiMo-V2.6-Pro has been recognized as the top-performing open-weights model on the Artificial Analysis Intelligence Index, offering strong cost efficiency.

The release of MiMo-V2.6-Pro by Xiaomi signifies a growing trend of established tech companies, beyond the typical AI-focused startups, entering the frontier AI model development space. Their approach, leveraging extensive RL compute and a focus on open-sourcing training infrastructure, suggests a potential shift in how cutting-edge models are developed and disseminated, challenging the dominance of closed, proprietary systems. This move, coupled with the ongoing advancements from other Chinese AI labs, intensifies competition and could lead to faster innovation cycles across the industry.

The strategic importance of open-source RL environments and training recipes is highlighted by Xiaomi's release, indicating that post-training optimization may be an increasingly cost-effective path to significant AI performance gains. This could democratize access to frontier-level capabilities, allowing smaller teams or researchers to achieve comparable results to larger, well-funded organizations. The focus on cost-efficiency and speed suggests a market leaning towards specialized, performant models rather than solely relying on massive, general-purpose ones, potentially reshaping the landscape of AI deployment and adoption.