GLM-4

This repository hosts the open GLM-4 releases from Zhipu AI (now Z.ai) and Tsinghua's THUDM group, with inference demos, fine-tuning scripts and model notes. The June 2024 wave brought GLM-4-9B with a 128K-context chat model, a 1M-context variant and the GLM-4V-9B vision model, all under a custom GLM-4 license. The April 2025 GLM-4-0414 wave reached 32 billion parameters: a base pretrained on about 15 trillion tokens, a chat model tuned for function calling, code and agent work, GLM-Z1 reasoning models at 32B and 9B, and a Rumination model that calls search tools while writing research-style reports. The 0414 models train at 32K context, with YaRN scaling beyond. Implementations exist in Transformers, vLLM and llama.cpp; the repo's own benchmarks show the 9B using about 19 GB in bf16 on short prompts and the 32B close to 70 GB, so quantized builds are the practical route on consumer cards. Repository code is Apache 2.0 and the 0414 weights are MIT, both permitting commercial use. Strong in Chinese and English, it suits teams wanting a permissive mid-size model; later releases like GLM-4.1V-Thinking have their own repositories.