GLM-4.5

GLM-4.5 is an open-source language model family for developers building AI agents and coding tools on their own servers. It supports reasoning and tool use, but its supplied inference configurations require server-class NVIDIA hardware.

• Reasoning modes: Choose thinking mode for complex reasoning and tool use, or non-thinking mode for direct responses.
• Tool calling: Build agents with vLLM or SGLang using OpenAI-style tool descriptions; both runtimes can separate reasoning output from the response.
• Model variants: Choose the full GLM-4.5 or the smaller GLM-4.5-Air, with base and hybrid reasoning models available in BF16 and reasoning models also available in FP8.
• Language and customization: Work in English and Chinese, and fine-tune the models through Llama Factory or Swift.

You can self-host inference with Transformers, vLLM or SGLang. The MIT license permits commercial use and modification. Z.ai and Zhipu AI also offer hosted APIs, where inference runs on their services rather than your hardware.

Hardware needs are substantial even for Air. The supplied FP8 inference configurations use eight H100 or four H200 GPUs for GLM-4.5, and two H100s or one H200 for Air. These configurations require more than 1 TB of server memory and assume FP8-capable hardware without CPU offloading. Using the full 128K context requires additional GPU capacity.