MiMo-V2.5-Pro

MiMo-V2.5-Pro is an open-source Mixture-of-Experts language model published by the Xiaomi MiMo team, with 1.02T total parameters and 42B active parameters. It inherits its hybrid attention design and 3-layer Multi-Token Prediction scheme from the earlier MiMo-V2-Flash, interleaving sliding window and global attention in roughly a 6:1 ratio with a 128-token sliding window and learnable attention sink bias to cut KV-cache storage by nearly 7x while preserving long-context behavior. Three lightweight MTP modules using dense feed-forward networks sit on top, roughly tripling output throughput during inference and helping accelerate rollout in reinforcement-learning training.

Positioned as the most capable entry in the MiMo line, MiMo-V2.5-Pro is aimed at agentic workflows, complex software engineering, and other long-horizon tasks, sustaining trajectories that span thousands of tool calls with coherent instruction following across a 1M-token context window. The open-weights release lets teams run the full model or build quantized derivatives, such as the FP4 variant discussed in the NVIDIA DGX Spark and GB10 developer community, making it a practical fit for self-hosted deployments that need strong reasoning, structured tool use, and extended context for multi-step engineering pipelines.