CogVLM2 Models by Tsinghua Knowledge Engineering Group (THUDM)
Last refreshed 2026-10-01. Next refresh: weekly.
Capabilities
About
CogVLM2 is THUDM's second-generation family of open vision-language models, following CogVLM, with weights on Hugging Face and ModelScope. The image models, cogvlm2-llama3-chat-19B (English) and cogvlm2-llama3-chinese-chat-19B (Chinese and English), are 19B models built on Meta-Llama-3-8B-Instruct for image understanding and multi-turn chat, with 8K text length and images up to 1344×1344. The first models were released in May 2024. CogVLM2-Video, released that July, understands videos of up to one minute by extracting keyframes, and comes in chat and base versions.
Decision facts
- Best fit
- vision and multimodal work
- Capability starting point
- CogVLM2 Chat 19B with multimodal inputs
- Lowest tracked input
- Not tracked
- Closest related family
- ChatGLM-4
Current Variants
Use-when guidance is based on each model's tracked capabilities, context window, release date, and replacement status.
Use when the workload needs 19B parameters and multimodal inputs.
| Model | Use when | Released | Signals | Status |
|---|---|---|---|---|
| CogVLM2 Chat 19B | Use when the workload needs 19B parameters and multimodal inputs. | 2024-05 | 19B parametersmultimodal inputs | Current |
Release Timeline
1 release groupSpecifications(1 models)
| Model | Released | Parameters | Multimodal |
|---|---|---|---|
| CogVLM2 Chat 19B | 2024-05 | 19B | Yes |
