Term
GLM-5
Open model from Zhipu AI (Z.AI), released in February 2026 under the MIT license. A mixture-of-experts with around 744B parameters (about 40B active per token), 200K context, focused on coding and agents. Freely downloadable via Hugging Face.
GLM-5 — explained in more detail
GLM-5 is an open language model from the Chinese lab Zhipu AI (also appearing as Z.AI), released in February 2026. It is available under the permissive MIT license on Hugging Face, making it free for open-source deployment and commercial use. GLM-5 is positioned as a powerful open model built to compete with leading proprietary models on reasoning, coding and agentic tasks.
Technically, GLM-5 is a mixture-of-experts (MoE) model with around 744 billion total parameters, of which only about 40 billion are active per token. The context window is 200,000 tokens. The architecture uses, among other things, a sparse attention mechanism and multi-token prediction, and emphasises programming and agent capabilities. In the reported benchmarks, GLM-5 achieves a high score on SWE-bench Verified (around 77.8%), among others.
Access is open: the weights can be downloaded and self-hosted; Zhipu AI also offers API access.
Example / practical relevance
GLM-5 is aimed especially at coding and agent scenarios: automated bug fixing in code repositories, multi-step developer workflows, or agents that use tools across many steps. The large 200K context makes it possible to process extensive code bases or documentation in a single pass.
Because the model is MIT-licensed, companies can run it themselves — attractive for data protection, cost control and independence from a single provider. The MoE architecture keeps inference costs comparatively low despite the very large total parameter count.
Distinction from similar terms
GLM is Zhipu AI’s model line and stands alongside other open families such as DeepSeek, Qwen (Alibaba) or Llama (Meta). GLM-5 is the fifth generation; later point versions (such as GLM-5.2, GLM-5.3) followed with further improvements. As an MoE model, GLM-5 differs from classic “dense” models where all parameters are always active.