88Z AI, 27 May — Kuaishou has officially open-sourced its self-developed multimodal large model, Keye-VL-2.0-30B-A3B. It introduces the DSA sparse attention mechanism to multimodal understanding for the first time, supporting a 256K ultra-long context window and enabling temporal causal reasoning for long videos. The model surpasses Gemini 2.5 Pro and Gemini 3 Flash on certain metrics in the TimeLens benchmark and unlocks Agent collaboration capabilities for complex tasks like coding, tool use, and search.
At a glance
- Introduces DSA sparse attention, supporting 256K context and long-video reasoning
- Outperforms Gemini series on some TimeLens metrics; features Agent collaboration
The update
Kuaishou has officially open-sourced its self-developed multimodal large model, Keye-VL-2.0-30B-A3B. This model introduces the DSA sparse attention mechanism to multimodal understanding scenarios for the first time. It supports an ultra-long context window of up to 256K tokens, enabling temporal causal reasoning for long-form video content. In terms of performance, Keye-VL-2.0-30B-A3B has surpassed Gemini 2.5 Pro and Gemini 3 Flash on certain metrics within the TimeLens video understanding benchmark. Furthermore, the model unlocks Agent collaboration capabilities for the first time, allowing it to handle complex task scenarios such as code generation, tool invocation, and information search, demonstrating strong potential for comprehensive applications.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- 量子位Source published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.