A model from academic research
Most of the large language models dominating the AI landscape come from industrial labs with billion-dollar budgets. GLM-4, developed by Tsinghua University in collaboration with Zhipu AI, demonstrates that academic research can produce competitive models. The project originates from Tsinghua’s Knowledge Engineering (KEG) research group, which has been working since 2020 on the General Language Model architecture — an approach that combines autoregressive and autoencoding objectives in a single pre-training framework.
The ChatGLM series has gone through several iterations. GLM-4 represents the generational leap: pre-training on a massive multilingual dataset, alignment via RLHF (Reinforcement Learning from Human Feedback), and capabilities that approach Western frontier models on standard benchmarks.
Architecture and variants
The GLM architecture differs from decoder-only models like GPT through its use of a pre-training objective based on blank infilling: the model learns to fill in missing text spans of variable length, combining bidirectional understanding with autoregressive generation. This architectural choice influences how the model handles context and produces text.
GLM-4 is released in 6B and 9B parameter variants, sizes that allow execution on consumer hardware with 16-24 GB of VRAM. The chat versions are optimised for multi-turn dialogue, with context windows up to 128K tokens in the extended variant.
Function calling and tools
One of GLM-4’s distinguishing features is native function calling support. The model can generate structured calls to external functions — APIs, databases, computation tools — following user-defined JSON schemas. This capability, already present in OpenAI’s models, is implemented in GLM-4 at the fine-tuning level, allowing integration of the model into application pipelines that require interaction with external systems.
Integrated code interpreter support allows the model to generate and execute Python code for solving mathematical problems or analysing data during a conversation.
Open licence and implications
GLM-4 is distributed under an open licence that permits commercial use. Model weights, inference code and fine-tuning instructions are available on GitHub and Hugging Face. For the global AI ecosystem, GLM-4 represents a reference point: it demonstrates that research groups outside the Silicon Valley circuit — OpenAI, Google, Anthropic, Meta — can produce models of comparable quality with fewer resources.
Link: github.com/THUDM/GLM-4