GLM

GLM (General Language Model) is a family of large language models developed jointly by Zhipu AI and Tsinghua University's Knowledge Engineering Group (THUDM). The series spans from the original open-source ChatGLM research models to a rapidly evolving commercial product line that in January 2026 made Zhipu AI — rebranded internationally as Z.ai — China's first major LLM company to complete a Hong Kong Stock Exchange IPO.

Dual identity: research origin and commercial product

GLM has always straddled academia and industry. The THUDM research group at Tsinghua University developed the foundational GLM architecture, and the ChatGLM-6B repository (41k+ GitHub stars) established GLM as a leading open-source bilingual model in the Chinese AI ecosystem. Zhipu AI commercialised the research into a product API and later rebranded to Z.ai internationally, while continuing to publish open-weight checkpoints under Apache 2.0.

Model families

• ChatGLM series (2023): The original open-weight bilingual dialogue models. ChatGLM-6B (6.2B params, 2K context) and ChatGLM2-6B (extended to 32K) established the series. ChatGLM3 added tool-calling and code capabilities with a 128K context variant.
• GLM-4 / GLM-4-9B (2024): The flagship generation introduced with the ChatGLM paper. GLM-4-9B offers 128K context (chat) and an experimental 1M context variant; GLM-4-9B-Chat is the primary open-weight deployment target. The GLM-4 All Tools variant integrates web browsing, Python code interpreter, and text-to-image generation — benchmarked against GPT-4 All Tools. Larger closed variants (GLM-4-Air, GLM-4-Plus) are API-only.
• GLM-4-32B-0414 (April 2025): A 32B parameter chat model, part of the April 2025 refresh alongside the Z1 reasoning line.
• GLM-Z1 (2025): The reasoning model branch. GLM-Z1-9B-0414 and GLM-Z1-32B-0414 are trained from the GLM-4 base using cold-start reinforcement learning targeting math, code, and logic. GLM-Z1-Rumination-32B-0414 uses extended deliberation — revisiting intermediate steps — for more open-ended complex tasks.
• GLM-4V-9B / CogVLM (vision): GLM-4V-9B is the open-weight vision-language model for image understanding. CogVLM (17B — 10B visual + 7B language, ~6.7k GitHub stars) is the multimodal research model for visual question answering, image captioning, and visual grounding.
• GLM-4.5 (July 2025): A Mixture-of-Experts model (355B total / 32B active parameters) targeting reasoning, coding, and agentic tasks. Released alongside GLM-4.5 Air (a smaller, cheaper variant). GLM-4.5-Flash is free on the API.
• GLM-4.6 (September 2025): A notable release built using Chinese domestic AI chips (Cambricon Technologies), demonstrating Zhipu's hardware independence from Nvidia.
• GLM-4.7 (2025): Subsequent iteration with 200K context window; available on Ollama (ollama run glm-4.7). GLM-4.7-Flash is free on the API.
• GLM-5 (February 2026): Major scale-up: 744B total / 44B active parameters (MoE, 256 experts with 8 active per token), 200K context window, pre-trained on 28.5T tokens. Benchmarked on SWE-bench and hard reasoning tasks.
• GLM-5.1 (April 2026, open-sourced): Formally known as GLM-Z1-Rumination, the open-weight reasoning model released in April 2026 that reportedly matched GPT-5.4 on coding benchmarks.

Access methods

Self-hosting: Open-weight models are published on Hugging Face under zai-org/ (formerly THUDM/). GLM-4-9B, GLM-Z1-9B/32B, GLM-4-32B, GLM-4V-9B, and GLM-5.1 can be run locally via Ollama (ollama run glm-4.7), vLLM, SGLang, llama.cpp/GGUF, and the Hugging Face transformers library. The 9B models fit on a single consumer GPU; the 32B models require 24GB+ VRAM. Weights are Apache 2.0 licensed for commercial use.

Zhipu AI API (open.bigmodel.cn / docs.z.ai): The OpenAI-compatible REST API provides access to all model tiers — open-weight and proprietary. Free tier: GLM-4.7-Flash and GLM-4.5-Flash are available at no cost to all registered users with no daily quota caps. Paid tier: GLM-4.7 and GLM-4.5 cost approximately $0.60/million input tokens and $2.20/million output tokens. Premium models (GLM-5, GLM-5.1, GLM-5-Turbo) are available via subscription plans. The API is primarily hosted in China; Z.ai also maintains international developer documentation at docs.z.ai.

Third-party: GLM-4.7 is available on OpenRouter; GLM models appear on select aggregator APIs including LiteLLM-compatible providers.

License: Apache 2.0 for all published open-weight models (GLM-4-9B, GLM-Z1 series, GLM-4-32B, GLM-4V-9B, GLM-5.1). Closed proprietary variants (GLM-4-Air, GLM-4-Plus, GLM-5, GLM-5-Turbo) are API-only.