CogVLM2 is an open-source multimodal large language model from Zhipu AI that the authors benchmark at GPT-4V level, built on Llama 3. This repository hosts the model code, weights and a video-capable sibling, CogVLM2-Video. It has around 2,400 GitHub stars.
CogVLM2 pairs a vision encoder with an 8B-parameter Llama-3 language model — roughly 19B parameters in total — and is trained to understand images at high resolution, read text in scenes, answer questions and hold conversations about visual content. On public multimodal benchmarks the authors report scores competitive with proprietary systems of its generation, which made CogVLM2 a popular self-hosting choice in 2024–25. The release lineup includes chat models for images and a video variant that samples frames to answer questions about clips, plus variant checkpoints including a visual grounding model. Weights are downloadable for local deployment, and online demos are provided. The code is Apache-2.0; the project comes from the team behind the GLM family of models and was last updated in 2025.
CogVLM2 is free to download and run: the code is Apache-2.0 and the weights are published on Hugging Face. Larger CogVLM models are also offered through Zhipu AI’s commercial platform if you prefer an API.
Developers and researchers who need a strong open vision-language model they can self-host for document understanding, visual QA and multimodal chat, and teams prototyping video-understanding features.