GLM-5.3's flagship weights are out

Z.ai’s flagship GLM-5.3 model went open-weight on Hugging Face, roughly two weeks after the company launched it API-only on August 14. Z.ai had said twice at launch that it would hold the weights back for safety evaluation and hardening first, a response to the model’s unexpectedly strong cybersecurity capability, before releasing them once that review finished. The smaller GLM-5.3-Flash variant already shipped under an MIT license on August 26.

GLM-5.3 reuses the same roughly 744-billion-parameter mixture-of-experts base as GLM-5.2, with the gains coming entirely from post-training, the same base a stealth model on OpenRouter was traced back to earlier this month before Z.ai confirmed it as an early GLM-5.3 checkpoint. Z.ai’s own evaluations put it at 84.5% on the CyberGym benchmark, ahead of Claude Mythos 5 and GPT-5.6 Sol, and it lifted Terminal-Bench 3.0 from 4.6% to 28.3% over the prior generation. The model carries a 1-million-token context window and a 128,000-token output limit.

What it means for you

The staged rollout, API first, weights held for a safety pass, then a public release, is a template worth watching from a Chinese open-weight lab: it’s a middle path between Meta’s fully-open releases and closed-API-only labs. If you build on open-weight coding or security tooling, GLM-5.3 is now a real self-hosting option at a capability tier that was API-exclusive two weeks ago, which puts more pricing pressure on hosted coding-agent products.