Zhipu AI Releases GLM-4: Close to GPT-4, With an App Store Risk

Originally published in Chinese on January 16, 2024. GLM-4 performance and agent capability Zhipu AI, a Chinese foundation model company, released GLM-4. Its overall performance is much improved over the previous generation and approaches GPT-4. GLM-4 also sharply improves agent capability. GLM-4 All Tools can understand and plan complex instructions on its own from the user's intent, and freely call a web browser, a Code Interpreter, and a multimodal text-to-image model to finish complex tasks. The GLMs feature for customizing personal agents went live at the same time, so users can create their own GLM agent with simple prompt instructions. First impressions and the comparison strategy I used it for a little over half an hour, and overall it feels like a clear step up from the last generation, with somewhat stronger logic. The official figures use GPT-4 as the reference frame. I have not run a more detailed comparison, so I will not judge them for now. From the dimensions of comparison shown at the launch event, Zhipu compared its own model's capability in considerable detail and set GPT-4 as the clear benchmark. That may be a good approach but not an absolute one. The strategy is not wrong, since you compare against whoev