CogVLM

CogVLM

Quick answer: CogVLM is Zhipu AI's 17B-parameter open-weight visual language model, notable for using a dedicated "visual expert" module within its attention layers rather than a simple projection, and a widely-cited early open baseline in MLLM benchmark research.

At a Glance

Where CogVLM leads

• Fully open weights, early strong open baseline widely cited across MLLM benchmark papers

Where it lags

• Superseded by newer open-weight families on most current benchmarks

Best for: open-weight comparison baseline in multimodal benchmark literature.

What CogVLM Is

CogVLM introduces a "visual expert" module — trainable visual-specific parameters added to each attention and feed-forward layer of a frozen LLM — allowing deep fusion of vision and language features rather than shallow projection, built on a Vicuna-7B backbone.

Specifications