GLM-5.2 is the open model that works as a general agent

Nathan Lambert argues that GLM-5.2, the open-weight model Z.ai (formerly Zhipu) released in mid-June under an MIT license, is the moment open models became genuinely usable as general agents. His claim is specific: it is the first open model that feels right inside a coding harness, doing the kind of long, tool-heavy agent work that until now belonged to closed labs. The weights are on Hugging Face, and the model leans on Z.ai's SLIME reinforcement learning framework, performing best in its maximum thinking-effort mode.

The comparisons he points to are what make the case. On Arena's agent leaderboard GLM-5.2 in its top mode matched Claude Opus 4.8 running without thinking, and on Design Arena it came out ahead of Claude Fable. Lambert also does the arithmetic on the gap: 204 days separate Opus 4.5 in November 2025 from GLM-5.2, which lines up with the six-to-nine month lag often cited between American and Chinese labs. That lag is now measured in months, not generations. You can read his full argument here.

Why it matters

If you have been waiting for an open model you can actually drop into an agent loop instead of a closed API, Lambert says this is the first one that clears that bar, which changes the build-versus-buy math for anyone who cares about cost, control, or keeping work off someone else's servers.