GLM-5.3 Open Weights Ship After the Cyber Safety Delay
Z.ai has published the full GLM-5.3 open weights on Hugging Face, making good on the release it delayed in August for safety work. The model is a 753B-parameter mixture of experts in FP8, claims the best open-weights coding results available, and ships with the same cyber capabilities that triggered the pause, now measured openly on public benchmarks.
On this page
The weights are out, and the cyber claims came with them
In August, Z.ai launched GLM-5.3 as a hosted service while holding the weights back, saying stronger-than-expected cyber capability required more safety evaluation. That hold is over. The full weights are now public on Hugging Face under the zai-org organization, with the repository created on August 25 and updated as recently as September 4. The model card states the thesis plainly: cyber capability "developed faster than we expected." [1] [3]
The artifact is large. GLM-5.3 is a 753B-parameter mixture of experts stored in FP8, roughly 755GB of safetensors across 141 shards, built on the same base as GLM-5.2 with every gain coming from post-training. The accompanying technical report, "GLM-5: from Vibe Coding to Agentic Engineering," is on arXiv. [1]
What the benchmarks actually show
Z.ai's headline claim is that GLM-5.3 is the most capable open-weights model for coding, with a 50 percent improvement over GLM-5.2 on the company's own code benchmark. The third-party-style numbers in the model card back a strong story: Terminal Bench 2.1 rose from 81.0 to 88.2, Terminal Bench 3.0 jumped from 4.6 to 28.3, and DeepSWE climbed from 46.2 to 66.9. The card claims open-source state of the art on Terminal Bench 3.0 and Agents' Last Exam, with evaluations run against Kimi K3, DeepSeek-V4 Pro, Qwen3.8-Max, Opus 4.8, Fable 5, and GPT-5.6 Sol. [1] [2]
The cyber numbers are the ones that caused the delay, and they are now published openly: CyberGym at 84.5 for vulnerability discovery, ExploitBench more than doubling from 24.4 to 54.4, and ExploitGym results rising from 29 and 39 to 105 and 130 on the two-hour and six-hour variants. Publishing these scores openly is itself notable: defenders can see exactly what offensive-capability curve they are planning around. [1]
What it takes to run it
Nobody is running 753B parameters on a phone, but the deployment picture is unusually good for a frontier-size model. The native checkpoint is FP8, and the community has already published 37 quantizations across llama.cpp, Ollama, LM Studio, and Jan. Official support covers SGLang, vLLM, Transformers, KTransformers, TokenSpeed, and Unsloth, plus Huawei Ascend NPUs through vLLM-Ascend and xLLM. Evals ran with 300K to 1M token contexts and up to 128K output tokens, and a reasoning_effort parameter defaults to max. [1]
For most people the practical route is hosted: the model is enabled on HuggingChat, sits behind six inference providers with output tokens starting around $4 per million, and had roughly 410,000 downloads at the time of writing. [1]
The license and safety questions
Two caveats belong in any coverage of this release. First, the license: unlike GLM-5.2 and the smaller GLM-5.3-Flash, a 320B MIT-licensed variant released on August 26, the flagship weights carry a custom glm-5.3 license rather than a standard open-source one. Check the terms before building a business on it. [1] [4]
Second, the safety question that caused the delay has no published answer. The model card describes benchmark hygiene, domain whitelists and LLM-based anti-cheat checks, but discloses no safety mitigations or evaluation outcomes from the pause itself. The weights shipped with their cyber capabilities intact, which is a defensible transparency choice and a real one: the same exploits these benchmarks measure are now in the hands of every defender and attacker who downloads them. Organizations running internet-facing services should assume models with this capability profile exist in the wild and patch accordingly. [1]
Sources
- zai-org/GLM-5.3 model cardHugging Face
- GLM-5.3: Frontier Coding with Emergent Cyber CapabilitiesZ.ai
- What is GLM-5.3Kie.ai
- zai-org/GLM-5.3-Flash model cardHugging Face