The year/Independent research

Paper 2606.10651

Kwai Keye-VL-2.0 Technical Report

Published
Jun 2026
Research lab
Independent
Citations
0
GitHub
809 stars

01 In brief

Summary

The report introduces Kwai Keye-VL-2.0-30B-A3B, an open-source Mixture-of-Experts (MoE) multimodal foundation model with 30B total parameters and 3B active, designed for long-video understanding and agentic intelligence.

It is the first to adapt DeepSeek Sparse Attention (DSA) to GQA-based architectures, enabling lossless 256K context processing.

The model uses a native-resolution vision encoder, unified visual encoding, and a four-stage pre-training curriculum (from projector initialization to 256K long-context extension).

Post-training includes supervised fine-tuning with synthetic CoT, reinforcement learning (General, Specialized, Video, Agentic RL), and Cross-Modal Multi-Teacher On-Policy Distillation (MOPD) to prevent catastrophic forgetting.

Efficient infrastructure includes ViT-LM heterogeneous parallelism, custom DSA kernels, and Chunk ViT inference.

Evaluations show state-of-the-art performance among similar-scale models, excelling in fine-grained temporal localization (TimeLens benchmarks) and long-video comprehension (Video-MME-v2, LongVideoBench), while remaining competitive on code, tool-use, and general vision-language tasks.

Model checkpoints are released to the community.

02 From the paper

Abstract

We introduce Kwai Keye-VL-2.0-30B-A3B, an open-source Mixture-of-Experts (MoE) multimodal foundation model designed to advance long-video understanding and agentic intelligence. To address the challenges of ultra-long contexts, information redundancy, and prohibitive computational costs inherent in hour-level videos, Keye-VL-2.0 is the first to adapt DeepSeek Sparse Attention (DSA) to GQA-based multimodal architectures, enabling lossless 256K context processing while capturing critical frames and long-range temporal dependencies. This architecture is underpinned by a highly optimized training and inference infrastructure, including scalable video I/O, heterogeneous ViT-LM parallelism, and custom DSA kernels that significantly maximize throughput and minimize computational overhead. Furthermore, to overcome the algorithmic dilemma of catastrophic forgetting during multi-task alignment, we introduce Cross-Modal Multi-Teacher On-Policy Distillation (MOPD) paired with Context-RL and Video-RL. By distilling dense token-level teacher feedback from on-policy rollouts back into the MoE backbone, which activates only 3B parameters, Keye-VL-2.0 natively empowers advanced agent collaboration across Code, Tool, and Search scenarios with multimodal self-correction. Extensive evaluations across video understanding, temporal grounding, reasoning, STEM, and agent benchmarks demonstrate that Keye-VL-2.0-30B-A3B achieves state-of-the-art performance among models of similar scale, particularly excelling in fine-grained temporal localization on TimeLens and long-video comprehension on Video-MME-v2 and LongVideoBench. We release our model checkpoints to accelerate community progress toward scalable and robust multimodal agentic applications.