GLM-5.3 tops open coding, but Z.ai is holding the weights back
All its lead comes from post-training, yet the download is gated over an unplanned cyber capability — leaving Coding-Plan access, no API, and no license yet.
Copy markdown
Top of the open coding board, on post-training alone
GLM-5.3 keeps GLM-5.2’s exact base model but jumps to 28.3 on Terminal Bench 3.0 (up from 4.6) and 66.9 on DeepSWE v1.1 — the strongest open-weight coding scores reported so far. Every gain is post-training; there’s no new pretrain to wait on.
The weights are held back — and the reason is the story
Z.ai is gating the download (the Hugging Face repo still 401s) until safety hardening finishes, reportedly around two weeks out, after the model showed an unplanned cyber capability that “outgrew its training.” For now there is no self-hosting an “open” model that you can’t actually pull.
How you actually reach it today
Right now GLM-5.3 is GLM Coding Plan–only: no standalone pay-per-token API and no published price. If you were planning to wire it into an agent via API or run it locally, budget for a wait rather than a weekend swap.
Specs, and the open-license caveat
1M-token context, 128K max output, and a community-estimated ~743B params (Z.ai hasn’t published the count). GLM-5.2 shipped under MIT, but 5.3’s license is still unannounced — don’t assume the same permissive terms until the weights and the license actually land.