Skip to main content
Back

Kuaishou's Keye-VL-2.0 Gets Smarter With More Video Frames, Beating Gemini

Kuaishou's new 30B mixture-of-experts vision-language model ships DeepSeek Sparse Attention to push 256K context and hour-long video understanding into open source.

  • Kuaishou released Keye-VL-2.0-30B-A3B, a 30B/3B-active MoE vision-language model under Apache 2.0.
  • First multimodal model in production with DeepSeek Sparse Attention, enabling 256K context windows.
  • Accuracy increases from 35.3% to 42.4% on VideoMME V2 as frame count scales from 64 to 512.
  • Beats Qwen3-VL-235B on LongVideoBench (74.1) and matches Gemini 3 Flash on TimeLens temporal grounding.
  • Ships with built-in Code, Tool, and Search agent capabilities plus visual self-correction.
  • Runs on SGLang, vLLM, and Transformers with OpenAI-compatible APIs.

Kuaishou's Keye team has dropped a new flagship multimodal model that does something unusual: it gets more accurate as you feed it more video frames, instead of degrading. Keye-VL-2.0-30B-A3B is a 30-billion-parameter mixture-of-experts vision-language model with 3B active parameters, released under Apache 2.0, and it claims to be the first multimodal system to ship DeepSeek Sparse Attention in production.

The model is aimed squarely at long-form video understanding and agent workflows, two areas where open-source VLMs have historically struggled against closed-source incumbents like Gemini. The pitch is that you can now do hour-long video reasoning with a 256K context window on a model you can actually download.

The 3B-active sweet spot

Keye-VL-2.0-30B-A3B is the latest 30B-class flagship base model in the Keye series, purpose-built to push the frontier of long-video understanding and to unlock the first generation of Agent capabilities in the Keye family. The A3B suffix refers to the active parameter count: of the 30B total weights, only about 3B are activated per token thanks to the mixture-of-experts routing. That makes inference roughly as cheap as running a small dense model while preserving the knowledge capacity of a much larger one.

Hugging Face lists it at 31B parameters in BF16 and F32 tensor types. For deployment, the team recommends SGLang on H800 GPUs with tensor parallelism of 2, though the model also works through vLLM and the standard Transformers pipeline with trust_remote_code=True.

Why DSA matters here

The headline architectural choice is DeepSeek Sparse Attention (DSA). Standard attention scales quadratically with sequence length, which is brutal when a single hour-long video can blow past hundreds of thousands of tokens. DSA replaces dense attention with a sparse pattern that only computes a fraction of the token-to-token interactions, letting the model handle very long sequences without the usual prefill cost explosion.