Xiaomi releases MiMo-v2.5 Family weights with strong coding and agent benchmarks (firethering.com)

Xiaomi has unveiled the MiMo-v2.5 Family, specifically the MiMo-V2.5-Pro model, which has showcased remarkable capabilities in coding and complex task execution. In a stunning demonstration, it completed a university-level compiler project in just 4.3 hours, achieving perfect scores on all hidden tests, outpacing traditional student efforts. The MiMo-V2.5-Pro boasts improvements over its predecessor, including enhanced long-horizon coherence and agentic capabilities, enabling it to handle over a thousand tool calls without losing focus. Notably, it significantly outperformed benchmarks in analog circuit design, showcasing its potential for demanding engineering tasks. This release is significant for the AI/ML community as it signals a leap in open-source coding models, with Xiaomi's model directly competing against established heavyweights like Claude Opus and GPT-5. Furthermore, MiMo-V2.5-Pro demonstrates improved token efficiency, reportedly using 40-60% fewer tokens than its competitors for comparable results. This efficiency can make a substantial difference for developers and researchers, particularly in agentic workflows that require extensive task completion over large contexts. Alongside the Pro version, Xiaomi has also introduced a multimodal variant capable of processing text, image, video, and audio, further broadening its applicability in diverse AI tasks.