Fugu-MT: arxivの論文翻訳
Daily Report / 2026-06-09

Daily Paper Briefing

2026-06-09 1 Paper

Kwai Keye-VL-2.0 Technical Report

Authors 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
Affiliations Kuaishou Group
Categories Method / Multimodal Models / Architecture combining video and agent intelligence, Method / Attention Mechanisms / Sparse attention in multimodal architecture, Method / Reinforcement Learning / Algorithms addressing catastrophic forgetting
License CC BY 4.0

Abstract Overview

Kwai Keye-VL-2.0-30B-A3B is an open-source multimodal Mixture-of-Experts foundation model aimed at long-video understanding and agentic intelligence. The report centers on two technical themes: extending multimodal context handling to 256K tokens with a GQA-compatible DeepSeek Sparse Attention design, and consolidating diverse post-training capabilities with Cross-Modal Multi-Teacher On-Policy Distillation alongside Context-RL and Video-RL. The system also uses native-resolution visual encoding, unified image/video processing, and optimized training and inference infrastructure for variable-length multimodal workloads. Evaluations cover long-video understanding, temporal grounding, code and tool-use agents, and general vision-language benchmarks.

Novelty

The paper claims to be the first to adapt DeepSeek Sparse Attention to a GQA-based multimodal architecture, enabling lossless 256K-context processing for long video. It also introduces Cross-Modal Multi-Teacher On-Policy Distillation to merge heterogeneous RL-trained teachers across text, image, video, and agent tasks while mitigating catastrophic forgetting.

Results

The model achieves the best reported results among the compared systems on LongVideoBench, Video-MME-v2 accuracy, all three TimeLens temporal-grounding subsets, tau^2-Bench, and VitaBench, while remaining competitive on MLVU, Video-MME, Video-MMMU, coding benchmarks, and broader vision-language evaluation. The report also states that Video RL improves general video benchmark performance by about 1 percentage point, and that the infrastructure yields substantial inference and training efficiency gains for long-context use.

Key Points

  1. Keye-VL-2.0 combines a 30B-class MoE backbone with only 3B active parameters, native-resolution visual encoding, and unified image/video tokenization to support detailed multimodal reasoning.
  2. Its long-context design uses a Lightning Indexer plus GQA sparse aggregation with top-k token selection, reducing attention complexity from O(L^2) to O(Lk) while supporting 256K multimodal contexts.
  3. Its post-training stack mixes SFT, synthetic CoT, General RL, specialized RL, Video RL, and MOPD to consolidate video, reasoning, coding, tool-use, OCR, grounding, and counting capabilities into one model.

References

This page was created using generative AI such as GPT-5, Claude Opus 4, Gemini 3, Gemini 3.1 Flash Image, and their higher-end successor versions. No guarantee can be made regarding its contents.