Z.AI ships GLM-5.2 under MIT, weights on Hugging Face

Z.ai released GLM-5.2 this week, and the part that matters for this crowd is the licence: it’s MIT, with the weights published on Hugging Face and ModelScope and day-one support across transformers, vLLM, SGLang, xLLM and ktransformers. No regional gate, no usage strings attached. After the month we’ve had watching frontier models get pulled behind export controls, “here are the weights, do what you like” is the refreshing end of the spectrum.

This is a 753B model. It’s data-centre-scale, not something you’re running on a homelab box without serious quantisation and a lot of VRAM. The “open” here is about control and study, not about it fitting on your desk.

Z.ai pitches it at long-horizon agentic work — large refactors, automated research, debugging sessions measured in hours. On their own benchmarks it posts 81.0 on Terminal-Bench 2.1 (against Claude Opus 4.8’s 85.0) and 62.1 on SWE-bench Pro, up from GLM-5.1’s 58.4. Vendor-reported numbers, so salt to taste until third parties run it.

Still: a permissively-licensed open-weight model trading blows with the closed frontier is the trend worth tracking. AI is getting really serious.

Sources

• GLM-5.2: Built for Long-Horizon Tasks — Z.ai on Hugging Face