ChatGLM3-6B

ChatGLM3-6B, published in late October 2023, closes the ChatGLM naming line - everything after it is branded GLM-4. Its contribution is the prompt format. This is the first model of the family with a redesigned prompt structure that natively carries function calls, a code interpreter and agent-style tasks alongside ordinary multi-turn dialogue, which put a six-billion-parameter local model in the same conversation as hosted assistants that had those features first. The base model was retrained rather than patched: more varied data, more training steps, a revised schedule. The vendor claimed the strongest performance among pre-trained models below 10 billion parameters at the time - a claim about a crowded field in late 2023 and worth reading as a snapshot, not a standing record. The shipped configuration is the detail to check. The default context is 8,192 tokens, not the 32,768 of ChatGLM2-6B; the long-context option is a separate release, ChatGLM3-6B-32K. The architecture is otherwise unchanged from the second generation: 28 layers, hidden size 4096, 32 attention heads over 2 key-value heads, a 65,024 token vocabulary, 6.24 billion parameters in float16. A base model, ChatGLM3-6B-Base, was released alongside it. Licensing follows the family pattern: a bespoke vendor document, free for academic research, free commercially only after registration through the vendor's form.

▸Videos

No videos yet.

▸Reviews