Xiaomi has released MiMo‑V2.5, an open‑weight AI model that the company says pushes the envelope in agentic capability and multimodal understanding.

Trained on 48 trillion tokens, the 310 B‑parameter version delivers best‑in‑class results on Xiaomi’s own agentic benchmarks and matches closed‑source rivals on image and video tasks.

The smaller V2.5 matches the larger V2.5‑Pro on coding challenges at half the cost, while a 1‑T parameter Pro variant offers 42 B active parameters. Running the model locally demands a high‑end GPU, but the model is also available via API or on Hugging Face.

Consumers and developers now have a new, powerful, and openly accessible AI tool that rivals major proprietary systems, potentially reshaping how businesses deploy multimodal solutions without licensing fees.