MiMo-V2.5-Pro is an open-source Mixture-of-Experts language model built around a 1.02T-parameter sparse architecture with 42B active parameters, a design that concentrates massive capacity into a manageable inference footprint. Its hybrid attention architecture interleaves Sliding Window Attention and Global Attention in a 6:1 ratio with a 128-token sliding window, dramatically cutting KV-cache storage to preserve long-context performance while keeping memory demands low. The model also carries three lightweight Multi-Token Prediction modules using dense feed-forward networks, a feature inherited from its MiMo-V2-Flash predecessor that triples output speed during inference and sets the stage for faster reinforcement learning rollout. The result is a model engineered to sustain complex trajectories—thousands of tool calls, deep instruction following, and multi-step reasoning—over an expansive context window.

On the benchmark stage, MiMo-V2.5-Pro tied for first among all open-weights models by scoring 54 on the Artificial Analysis Intelligence Index, putting it in the same tier as the most capable frontier systems. It holds top rankings on ClawEval, GDPVal, and SWE-bench Pro, and posts a strong 73.7 on SWE-bench Pro—numbers that highlight its particular strength in software engineering and agentic claw tasks. Released under the permissive MIT License, both the base and instruction-tuned variants are freely downloadable from HuggingFace, making them ready for local deployment or private-cloud hosting in commercial products. The model family matches frontier-level benchmarks while operating at a fraction of the typical token cost, an efficiency advantage that makes it especially attractive for high-volume agentic deployments where cost per task compounds at scale.