Reproduce “Kimi-VL Technical Report” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Kimi Team, Angang Du, Bohong Yin, Bowei Xing, Bowen Qu, Bowen Wang, Cheng Chen, Chenlin Zhang and 87 more

Kimi-VL is an efficient open-source Mixture-of-Experts vision-language model that achieves competitive performance across diverse multimodal tasks, often matching or exceeding larger proprietary and open-source models with only 2.8 billion activated parameters. Its "Thinking" variants enhance reasoning, while the architecture supports 128K extended context and native-resolution visual inputs.

Key insights

  • High performance in vision-language models is achievable with an efficient Mixture-of-Experts (MoE) architecture, significantly reducing activated parameters.
  • Combining native-resolution vision encoding with an extended context window (128K) is crucial for comprehensive understanding of high-detail and long-form multimodal content.
  • Long Chain-of-Thought (CoT) supervised fine-tuning and reinforcement learning effectively imbue models with robust multi-step reasoning and agent capabilities in multimodal settings.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.