GLM-5.2 is the best open-weight model yet, and you can't run it at home

Z.ai dropped GLM-5.2 on June 13 with the weights under an MIT license, and on paper it is the best open model anyone has shipped. It beats GPT-5.5 on SWE-bench Pro, 62.1 to 58.6, at roughly a sixth of the per-token cost, and it posts the highest score of any open-weight model on the Artificial Analysis intelligence index. If you write code against a model all day, those are not small numbers.

Then you look at what it takes to run.

GLM-5.2 is a 744-billion-parameter mixture-of-experts model. Crush it down to a brutal 2-bit quant and you are still staring at something like 245GB of memory just to load the thing. That is not a 24GB card and an optimistic quant. That is a server, or a rented cloud GPU, and for almost everyone reading this it is neither.

So here is the honest version, because this is the whole reason Will It Run exists. “Open weights” and “runs on your hardware” are two different sentences, and GLM-5.2 is where the gap turns into a canyon. The MIT license is real and it matters: it feeds the distills and the smaller GLM variants that will actually land on desks over the next few months. It does not put the flagship next to you. Anyone selling you “frontier coding, local, today” on the back of this release is quietly selling you a cloud bill.

If you have an API budget and you write software, GLM-5.2 at a sixth of GPT-5.5’s price earns a real look this week. If you wanted it on the bench, watch the memory math instead of the benchmark chart, because that is the number that decides whether “open” means anything to you. Open weights you cannot run are just someone else’s server with better marketing.