Official Z.ai platform to experience our new, MIT-licensed GLM models (Base, Reasoning, Rumination). Simple UI focuses on model interaction. Free.
This is the 12th launch from Z.ai. View more
GLM-5.2
Launching today
GLM-5.2 is Z.ai's flagship coding model with a usable 1M token context window and two reasoning effort levels. Open weights coming under MIT license.
For developers building long-horizon coding agents and large codebase refactors.
Free Options
Launch Team
Hugging Face
Z.ai just launched GLM-5.2, the third major release in its GLM-5 line, positioned specifically for long-horizon coding and agentic tasks.
The core problem this targets is simple to describe and hard to fix: a model can technically support a huge context window, but if it can't stay coherent across that whole window, the extra tokens don't help. GLM-5.2's solution is specialized training focused on making the full 1M token context stable enough for sustained plan, execute, test, fix loops rather than just single-turn answers.
The two reasoning modes (max for ceiling performance, high for token efficiency) paired with a 1M token context mean you get a model that holds full architectural context across a long session, so multi-hour autonomous coding runs are less likely to drift off track or violate constraints set early on.
Z.ai reports GLM-5.2 outperforms GPT-5.5 and Opus 4.7 on multiple benchmarks while trailing Claude Opus 4.8 by just 1% on FrontierSWE, which if it holds up under independent testing, means meaningfully fewer manual corrections during long agentic coding sessions.
It's aimed at engineering teams already running coding agents on large repositories who've specifically hit context-reliability walls with other long-context models.
Important to know upfront: access right now is limited to the GLM Coding Plan tiers. The open weights, standalone API, and public chatbot are all promised for next week, so this is a preview rather than a fully open release just yet.
If long-context reliability has been your bottleneck with coding agents, this is one to watch closely as the open access rolls out. Try it through the Coding Plan now, or wait a week for the open weights.