Kwai Publishes the Keye-VL-2.0 Technical Report on Long-Video Understanding
Kwai Keye published arXiv:2606.10651 on June 10, 2026 after releasing Apache 2.0 weights on Hugging Face on May 25. The vision-language model has 30B total and 3B active parameters, a 256K-token context, and DeepSeek Sparse Attention for hour-scale video processing. The team reported gains over open-source competitors on five video benchmarks and performance matching or exceeding Gemini 3 Flash. It also supports coordination between search, tool, and coding agents.
Visit original source