glm-5-2

Overview. GLM-5.2 is Z.ai's (formerly Zhipu AI) flagship open-weights reasoning and agentic-coding model, built as a 744B-total-parameter mixture-of-experts network with 40B active parameters per token. It ships under an MIT licence with a 1,048,576-token context window, making it a fit for developers who need very long-context reasoning or agentic coding workflows and want the option to self-host the weights rather than depend solely on a hosted endpoint.

Key strengths. Artificial Analysis has named GLM-5.2 the leading open-weights model on its Intelligence Index, and its domain profile skews heavily toward code, consistent with Z.ai's own framing of the model as an agentic-coding flagship. It runs as a reasoning model with selectable effort levels (high and xhigh), giving developers a visible reasoning trace they can tune for harder multi-step problems. The 1M-token context window is large enough to hold sizeable codebases or long agent transcripts in a single call.

When to use it. Reach for GLM-5.2 on complex, multi-step coding or agentic tasks where you want an extended reasoning trace and can tolerate slower, batch-style turnaround rather than snappy chat latency. It's a strong fit when a task needs to reference very large amounts of context in one pass — think large repositories, long documents, or extended tool-use transcripts — and where self-hosting the open weights under MIT terms is valuable for control over the inference environment.

Limitations. Output is capped at 128K tokens per response, and the model is text-only — there's no vision or audio input or output support. Its reasoning-driven response style falls into a batch latency class, so it isn't the right choice for latency-sensitive interactive chat. No official training-data cutoff date has been published by Z.ai.

Sovereignty note. GLM-5.2 is created by Z.ai (Zhipu AI), a China-based developer, not by Mistral as currently recorded in this catalogue entry — that attribution needs correcting before publishing, since it changes the country-of-origin and licensing-policy facts shown to customers. Because the weights are openly licensed under MIT, the model can be self-hosted in an environment of your choosing for full control over where inference data resides, independent of any hosted provider's policies.