Kuaishou Open-Sources Keye-VL-2.0-30B-A3B — Apache-2.0 Mixture-of-Experts Vision-Language Model Activates 3B Parameters Per Token, Lands 256K Context With DeepSeek Sparse Attention for Lossless Long-Video Reasoning, Beats Qwen3-VL-235B on LongVideoBench at 74.1, and Tops 70.1 mIoU on QVHighlights-TimeLens

What is it?

Keye-VL-2.0-30B-A3B is Kwai (Kuaishou)'s flagship multimodal foundation model in the Keye family. It is a sparse Mixture-of-Experts with 30B total parameters but only ~3B active per token, paired with a vision encoder for image and hour-long video input. It is the first multimodal model to land DeepSeek Sparse Attention (DSA) in production, which is what unlocks the 256K context.

How does it work?

The model combines DSA with GQA so token-by-token compute stays close to a 3B dense model while the full 30B parameter budget is available to route through. Training uses Cross-Modal Multi-Teacher On-Policy Distillation to fight catastrophic forgetting during multi-task alignment. A built-in Agent collaboration scaffolding wires Search, Tool, and Code paths into the base model rather than bolting them on later. Sparse attention plus targeted feature aggregation lets it scan hours of video frames without losing temporal grounding.

Why does it matter?

Open-weights long-video understanding has been bottlenecked by both compute and benchmarks. Keye-VL-2.0 beats Qwen3-VL-235B on LongVideoBench at 74.1 with one-eighth the active parameters, posts 70.1 mIoU on QVHighlights-TimeLens, and ships under Apache-2.0 with code, model weights, and a public demo. Combined with a permissive license, that turns it into a credible drop-in for long-video agent stacks built on closed APIs.

Who is it for?

open-weights ML engineers, video understanding researchers, agent builders

Try it

huggingface-cli download Kwai-Keye/Keye-VL-2.0-30B-A3B

Tags

• kuaishou
• kwai-keye
• vision-language
• mixture-of-experts
• deepseek-sparse-attention
• long-video
• agent
• multimodal
• apache-2.0
• open-weights