Kwai Keye-VL-2.0 Technical Report

Benchmark Model Rank Results
agents-on-benchKeye-VL-2.0-30B-A3B#5Accuracy: 82.6
temporal-sentence-grounding-on-activitynet-timelensKwai Keye-VL-2.0-30B-A3B#5mIoU: 58.5
temporal-sentence-grounding-on-charades-timelensKwai Keye-VL-2.0-30B-A3B#5mIoU: 58.4
agents-on-bfclv4Keye-VL-2.0-30B-A3B#6Accuracy: 65.7
zero-shot-video-question-answer-on-video-mme-w-o-subsKeye-VL-2.0-30B-A3B#8Accuracy (%): 78.3
video-question-answering-on-mlvuKeye-VL-2.0-30B-A3B#10Accuracy: 82.8
video-question-answering-on-video-mmmuKeye-VL-2.0-30B-A3B#12Accuracy: 80.0
multi-task-language-understanding-on-c-evalKeye-VL-2.0-30B-A3B#15Accuracy: 89.2
multimodal-reasoning-on-mmmuKeye-VL-2.0-30B-A3B#15Accuracy: 75.3
visual-grounding-on-refcoco-refcoco-refcocog-overallKeye-VL-2.0-30B-A3B#15Accuracy: 88.1
optical-character-recognition-ocr-on-omnidocbench-v1-5Keye-VL-2.0-30B-A3B#19Accuracy: 84.7
vision-language-understanding-on-mmstarKeye-VL-2.0-30B-A3B#19Accuracy: 75.9
multimodal-reasoning-on-mathvistaKeye-VL-2.0-30B-A3B#20Accuracy: 82.9
question-answering-on-humanity-s-last-exam-hleKeye-VL-2.0-30B-A3B (w/ CoT)#23Accuracy: 22.9
mathematical-reasoning-on-aime-2025Keye-VL-2.0-30B-A3B#29Accuracy: 86.7
optical-character-recognition-ocr-on-ocrbenchKeye-VL-2.0-30B-A3B#33Score: 857
code-generation-on-livecodebench-v6Keye-VL-2.0-30B-A3B#40Pass@1: 64.2
code-generation-on-swe-bench-verifiedKeye-VL-2.0-30B-A3B#59Accuracy: 62.0