Xiaomi has released the MiMo model family, featuring a flagship version with 1.02 trillion parameters. The company published the model weights under the MIT license. The MiMo lineup includes specialized models for reasoning, vision, audio, and robotics, developed over an 18-month period.
The flagship MiMo-V2.5-Pro utilizes a sparse mixture-of-experts architecture with 384 experts, selecting eight per token. It supports a context window of up to 1,048,576 tokens. To optimize performance, 60 out of 70 layers focus on the 128 most recent tokens, while only 10 layers process the full context.
The UltraSpeed configuration achieves over 1,000 tokens per second on a single eight-GPU node. This is enabled by compressing expert weights to MXFP4 format, speculative decoding, and TileRT kernels. The total weight size for the trillion-parameter model is approximately 510 GB.
Post-training for the models relies on the MOPD method, where specialist models provide token-level hints to the student model. Additionally, the R3 routing mechanism ensures consistency between inference and training phases. Xiaomi discontinued hosting for the V2 version on June 30, 2026.
