Xiaomi Releases MiMo-V2.6-Pro 1T-A42B Open-Weights Model Trained for $3M
Xiaomi has released MiMo-V2.6-Pro 1T-A42B, a top-performing open-weights large language model featuring a 1-trillion parameter architecture. Remarkably, the company achieved frontier-level model capabilities with an efficient training budget of just $3 million. This launch establishes Xiaomi as a major frontier AI lab while demonstrating extreme cost efficiency in training massive models. It offers the open-source community access to a highly capable frontier LLM, lowering the barrier for high-performance AI research and deployment. The model's '1T-A42B' designation indicates a total capacity of 1 trillion parameters with 42 billion active parameters per token, characteristic of a Mixture-of-Experts (MoE) architecture. The $3 million training cost highlights significant algorithm and compute infrastructure optimizations by Xiaomi's AI team.
## BACKGROUND
Open-weights models allow developers and researchers to freely download model parameters and run them on custom hardware. Mixture-of-Experts (MoE) architectures scale total parameter counts massively while keeping inference costs low by routing each token to only a sub-network of active parameters.