GLM-5.2 is a large-scale open-weights reasoning model built around a Mixture-of-Experts architecture with 753B total parameters and roughly 40B active at inference, released under an MIT license after a mid-June 2026 rollout. The model is text-only on both input and output, with vision handled by a separate model family rather than bundled in. It is paired with an extended context window suited to project-scale work, and the provider positions it for long-horizon agent workflows, multi-step automation, and end-to-end software engineering tasks that need consistent engineering context from requirements through deployment.

In third-party and community testing, GLM-5.2 has drawn attention for coding and tool use, with an independent benchmark write-up reporting it ahead of a leading closed competitor and Hacker News anecdotes describing fast output and workable behavior when wired into real agent pipelines. OpenRouter exposes reasoning effort controls including high and xhigh modes for heavier deliberation, and the model is routable across multiple host providers at modest token pricing, which broadens practical access. The combination of open weights, a large MoE design, and agent-oriented strengths makes GLM-5.2 a natural fit for teams building coding assistants, autonomous tooling, and other long-running workflows where sustained reasoning and tool use matter more than multimodal breadth.