Zhipu open-sources ChatGLM-6B

Zhipu AI, a spin-off of Tsinghua University’s Knowledge Engineering Group, released ChatGLM-6B, an open-weight, bilingual conversational model based on its General Language Model architecture. At 6.2 billion parameters, it was a much smaller sibling of Zhipu’s larger GLM-130B, aligned for dialogue through supervised fine-tuning on human-curated conversation data followed by reinforcement learning from human feedback.

The release’s distinguishing feature was accessibility. Trained on roughly a trillion tokens of mixed Chinese and English text, the model could be run with as little as 6GB of GPU memory at INT4 quantisation — within reach of a single consumer graphics card, rather than the multi-GPU server clusters that running most contemporary chat models required. The weights were made freely available for academic research, with free commercial use permitted after registration, under a licence for the code itself of Apache 2.0.

ChatGLM-6B arrived amid a wave of interest in ChatGPT-style assistants following OpenAI’s November 2022 launch, at a point when very few competitive open-weight conversational models of any size existed publicly, let alone ones capable of fluent Chinese. It attracted rapid adoption within the Chinese open-source AI community and became one of the most widely used bilingual open models of 2023, predating the larger open-weight releases from Alibaba’s Qwen and DeepSeek that followed over the next two years and came to dominate that space.