GLM-5.3 is now open-weight

	GLM-5.3 is now open-weight(huggingface.co)
	797 points by jeudesprits 13 days ago \| 279 comments
	tl;dr: Z.ai released GLM-5.3 as open-weights, built on the same base as GLM-5.2 but with post-training improvements yielding ~50% coding gains and open-source SOTA on Terminal Bench 3.0 and Agents' Last Exam. The model also shows sharp jumps in cybersecurity capability, more than doubling GLM-5.2 on exploitation benchmarks and topping CyberGym. It supports deployment via SGLang, vLLM, Transformers, and Ascend NPUs, with a configurable reasoning_effort parameter (low/high/max).
	HN Discussion: ↑GLM-5.3 is a sweet spot open-weights model, easier to run and less restrictive than US alternatives ↑GLM-5.3 shows impressive capability, tackling hard problems better than competing models like DS4Flash ~Chinese models including GLM tend to overthink and rack up token costs despite cheaper per-token pricing ↓GLM-5.3 Flash is more interesting since 5.3 lacks vision support, which is a dealbreaker ↓The release doesn't mean much yet due to limited architecture support and infrastructure availability