Xiaomi releases MiMo-V2.6, a trillion-parameter open model, and details training costs

Xiaomi Debuts MiMo-V2.6 Open Model

Xiaomi has introduced MiMo-V2.6, an open-weight AI model, including the flagship MiMo-V2.6-Pro. This model features a trillion parameters, with 42 billion active at any given time, and a one-million-token context window. It supports multimodal inputs, including text, images, audio, and video, placing it in a similar category to frontier models from other major AI developers.

Performance and Independent Validation

Xiaomi claims frontier-level performance for MiMo-V2.6 across various tasks such as coding, agentic functions, cybersecurity, multimodal work, and research. Independent analysis by Artificial Analysis supports these claims, giving MiMo-V2.6-Pro an Intelligence Index score of 46, which positions it as the top-ranked model among the 114 large open-weight models tracked by the firm.

Unprecedented Training Transparency

A notable aspect of Xiaomi's release is its transparency regarding the model's training. The company livestreamed its reinforcement learning (RL) training process through a public dashboard over five days, starting September 15. This public record displayed real-time metrics from the production RL runs.

Disclosure of Training Costs

During the livestreamed training, Xiaomi disclosed the costs associated with the RL phase. The dashboard showed expenses of $854,044 for the smaller MiMo-V2.6-Flash model and $2,620,670 for the Pro version, totaling approximately $3.5 million. This public disclosure of RL training costs is uncommon in the industry, with previous instances from MiniMax and DeepSeek providing similar, though lower, figures for their respective models.