What's this about?
Four days after releasing GLM-5.3-Flash, Z.ai open-weighted the full GLM-5.3. Their own description is worth reading twice: "our most capable model for agentic coding and cyber defense is now available to download, run, and customize." It ships as 141 shards, roughly 770GB unquantized — about the same size as GLM-5.2, which several commenters note means the gains came from training, not scale.
What does "open-weight" actually give you?
The trained parameters, downloadable, runnable on your own hardware, modifiable. It is not the same as open source — you generally do not get the training data or the training code, so you cannot reproduce it, only use and adapt it. The practical consequence is that nobody can revoke it. Once weights are public they stay public, which is exactly why the capability disclosure question below matters.
What's the community saying?
- The headline endorsement: a daily user says it "feels like Opus 4.8, in the best possible way" — running it through the z.ai coder plan in opencode and via direct API for real side projects.
- A sharp exchange follows about what that implies for Opus 4.8's size, with the sensible correction that you cannot compare models released six months apart, and that GLM 5.2 shared this architecture while being much worse.
- The size datapoint that grounds it: full precision unquantized is about 770GB, roughly unchanged from 5.2 — so this is a training-quality jump, not a parameter-count jump.
- Mollick's note runs alongside: as open models get this good, publishing model cards and doing red teaming stops being optional, because guardrails on any open model can be removed.