On May 9, Xiaomi officially open-sourced MiMo-V2.5-Pro — a Mixture-of-Experts (MoE) large model with 1.02T total parameters and 42B activated parameters, based on a Hybrid Attention architecture, with a context window of 1 million tokens.
The core breakthrough lies in long-horizon consistency. Officially disclosed tests show that in complex software engineering tasks requiring over a thousand Tool Calls (Peking University compiler principles course project: implementing a complete SysY compiler in Rust from scratch), MiMo-V2.5-Pro completed 672 tool calls in 4.3 hours, scoring 233/233, perfectly passing all hidden test cases. This isn't a routine benchmark score, but a real long-horizon autonomous task — the model needs to continuously self-correct, plan across stages, and any logical flaw in any step would cause final failure.
Architecturally, V2.5-Pro uses a Hybrid Attention mechanism, mixing standard Transformer's self-attention with linear attention, controlling compute complexity while preserving global modeling capability. As an MoE model, only 42B of the 1T total parameters are activated per inference, combined with a 1M context window, making per-request compute cost far lower than equivalent-scale Dense models.
Xiaomi simultaneously opened Hugging Face model weights and API interfaces, letting developers call them directly. Compared to closed large models that often require thousands of GPU-hours, MiMo-V2.5-Pro lets resource-constrained teams experience frontier Agent capability. This isn't just a model performance improvement, but a sign of the open-source ecosystem moving toward a truly usable stage.