MiMo-V2.5

MiMo-V2.5 is positioned as a major step toward agentic and multimodal understanding, built as a sparse mixture-of-experts model with 310B total parameters and 15B active, trained on 48 trillion tokens. Its language backbone inherits the hybrid sliding-window attention design from MiMo-V2-Flash, then gains native visual and audio understanding through in-house encoders joined by lightweight projectors. Training unfolds across five stages, from text pre-training and projector warmup through multimodal pre-training, supervised fine-tuning, agentic post-training that progressively widens the context window from 32K to 256K and onward to 1M tokens, and finally reinforcement learning plus an MOPD stage that strengthens perception, reasoning, and acting on what the model perceives.

On agentic benchmarks relevant to real deployment, MiMo-V2.5 delivers best-in-class scores against peer open models such as MiMo-V2-Pro, Kimi K2.6, and DeepSeek-V4-Flash, while also holding its own against frontier proprietary systems like Claude Opus 4.6, Gemini 3.1 Pro, and GPT-5.4. Because weights are openly available, the model can be self-hosted and quantized, and community runs have already demonstrated NVFP4 deployment on two DGX Spark units with vLLM, including full multimodal serving. The combination of long-context handling, native cross-modal reasoning, and open availability makes it a practical fit for teams building coding agents, document and media analysis tools, and other applications that need a single model that can see, hear, read, and act.