Xiaomi

Lightweight, high-speed reasoning model with hybrid attention and multi-token prediction for low-cost inference and strong benchmark scores.

Multimodal model with native visual and audio understanding on a 1M context, designed to reason and act across modalities in agentic workflows.

Top-tier model for agentic workflows, complex software engineering, and long-horizon tasks, sustaining work across 1000+ tool calls on 1M context.