GLM-4-9B-Chat

GLM-4-9B-Chat is the open-weights member of the GLM-4 generation, published in June 2024 and aimed squarely at Llama-3-8B, which the vendor named as the model it claimed to beat across semantics, mathematics, reasoning, code and knowledge tests. The practical jump over the ChatGLM generation is in what the model does rather than what it scores. This is the first model in the line with web browsing, code execution and custom function calls as advertised features, a 128K context window in the shipped configuration, and support for 26 languages rather than just Chinese and English. Two siblings shipped at the same time: GLM-4-9B-Chat-1M, which extends the window to a million tokens - roughly two million Chinese characters - and GLM-4V-9B, which adds vision at 1120x1120 resolution. Architecturally it stays inside the family: 40 layers, hidden size 4096, 32 attention heads over 2 key-value heads, a 151,552 token vocabulary, 9.4 billion parameters in bfloat16. The licence tightened compared with ChatGLM2. It is still a bespoke vendor document with a revocable grant and a commercial registration form, but it adds obligations that carry into derivative work: distributing anything built on GLM-4 requires shipping a copy of the agreement, displaying the notice 'Built with glm-4' on the product, and prefixing any derived model's name with 'glm-4'. Teams that fine-tune open weights and rebrand the result need to read that clause before shipping.

▸Videos

No videos yet.

▸Reviews