Xiaomi Launches MiMo-V2.5 with Enhanced Agentic and Multimodal Capabilities

Xiaomi has released MiMo-V2.5, a 310 billion-parameter Sparse MoE model that excels in agentic capability and multimodal understanding with native visual and audio comprehension. It supports up to 1 million tokens of context and delivers best-in-class performance on agentic benchmarks like Claw-Eval where it achieved 62.3. The model is also fully open-sourced, with weights and tokenizers available on Hugging Face.