Kwai Keye-VL-2.0 Technical Report
Kwai Keye Team, Bin Wen, Changyi Liu, Chengru Song, Chongling Rao, Guowang Zhang, Han Li, Haonan Fan, Hengrui Ju, Jiankang Chen, Jiapeng Chen, Jiawei Yuan, Kaixuan Yang, Kaiyu Jiang, Kun Gai, Lingzhi Zhou, Na Nie, Sen Na, Tianke Zhang, Tingting Gao, Xuanyu Zheng, Yulong Chen, Fan Yang, Haixuan Gao, Lele Yang, Mingqiao Liu, Muxi Diao, Qi Zhang, Qile Su, Wei Chen, Wentao Hong, Xingyu Lu, Yancheng Long, Yankai Yang, Yingxin Li, Yiyang Fan, Yu Xia, Yuzhe Chen, Ziliang Lai, Chuan Yi, Haonan Jia, Tianming Liang, Weixin Xu, Xiaoxiao Ma, Yang Tian, Yufei Han, Feng Han, Hang Li, Jing Wang, Jinghui Jia, Junmin Chen, Junyu Shi, Ruilin Zhang
arXiv:2606.10651 · 192▲ · cs.CV
View on arXiv →Premium readers get an interactive explainer for this paper — a figure you can poke at, not just read.
What it is
A technical report on Kwai Keye-VL-2.0-30B-A3B, an open-source Mixture-of-Experts multimodal model (30B total, 3B active) that adapts DeepSeek Sparse Attention (DSA) to a GQA-based backbone so it can process 256K-token contexts, aimed at hour-long video understanding. It also describes a post-training recipe combining Cross-Modal Multi-Teacher On-Policy Distillation (MOPD) with Context-RL and Video-RL to add agent, code, and search skills without wiping out base reasoning.
Why it matters
The main practical angle is inference cost for long video: DSA replaces dense O(L^2) attention with top-k token selection (k=2048), turning attention cost into O(Lk) and bounding KV cache growth, so you can feed hour-scale video without aggressive frame subsampling. For practitioners this means long-video comprehension becomes deployable on a model that only activates 3B parameters per token, though the claims rest on the team's own benchmarks and released checkpoints rather than independent replication.
Practical takeaway
If you work with long-video or multi-page document pipelines, watch for the released checkpoints as an open-source option that keeps native-resolution encoding and 256K context; the DSA-on-GQA approach is also a reference pattern if you are extending existing grouped-query attention models to long context rather than starting from MLA.
Key result
The report claims state-of-the-art results among similar-scale models on fine-grained temporal localization (TimeLens: ActivityNet, QVHighlights, Charades) and long-video benchmarks (Video-MME-v2, LongVideoBench), and reports top scores on tau^2-Bench (82.6) and VitaBench (33.1) for agentic tool use, with second place on BFCL-V4 (65.7). All figures are self-reported and several comparison columns in the tables are blank, so cross-model comparison is incomplete.
Subscribe
Get the next issue.
Free. One email a week. Unsubscribe any time: no account, no dark patterns.