Keye-VL-2.0-30B Model Released

The latest flagship model in the Keye series, Keye-VL-2.0-30B-A3B, is designed for long-video understanding and implementing AI agent capabilities.

Key features are as follows:

• Video Understanding and Spatiotemporal Localization: It overwhelms open-source competitor models across 5 video benchmarks, with temporal grounding performance matching or exceeding Gemini-1.5-Flash.
• DSA (DeepSeek Sparse Attention) Architecture: Through a sparse attention mechanism, it supports efficient computation and precise video understanding even in 256K ultra-long contexts.
• Agentic Capabilities: Built-in agentic collaboration mechanisms for code writing, tool use, and web search enable it to perform complex workflows.
• High-Efficiency Inference Stack: The DSA-Native architecture and various optimization techniques reduce prefill costs for long sequences and increase training throughput.