ChatGLM-6B

ChatGLM-6B was published in March 2023, three months after ChatGPT reached the general public and four months before Llama 2. It matters less for what it could do than for where it could run: 6.2 billion parameters, quantised to INT4, fit into 6 GB of graphics memory - a consumer card, not a data centre. For a large part of 2023 it was the default way to try a Chinese-language assistant on your own machine. The architecture is the GLM framework developed at Tsinghua University rather than a plain decoder transformer: 28 layers, hidden size 4096, 32 attention heads and a 2048-token sequence limit, trained on roughly a trillion bilingual tokens and aligned with supervised fine-tuning, feedback bootstrapping and reinforcement learning from human feedback. The licence is the part worth reading before use, and it has not changed since. The weights are released under a bespoke document, the ChatGLM-6B License, not an open-source licence: the copyright grant is explicitly revocable, military use is prohibited, commercial use requires filling in a registration form with the vendor, and disputes go to the People's Court of Haidian District in Beijing under the law of the People's Republic of China. The same document states that it covers this model only and not ChatGLM2-6B or anything after it. The repository is still up and still collects downloads, though the traffic long ago moved to its successors. It carries more likes than any other model on the company's account - a monument as much as a tool.

▸Videos

No videos yet.

▸Reviews