Kuaishou open-sources its self-developed multimodal large model Keye-VL-2.0-30B-A3B

Kuaishou has officially open-sourced its multimodal large-scale model, Keye-VL-2.0-30B-A3B, which for the first time introduces the DSA sparse attention mechanism into multimodal understanding scenarios. It supports ultra-long contexts of up to 256K bytes, enabling temporal causal inference in long videos. The model outperforms Gemini 2.5 Pro and Gemini 3 Flash in some metrics of the TimeLens video understanding benchmark, and for the first time unlocks Agent collaboration capabilities, covering complex task scenarios such as code, tool calls, and searches.