Z.ai opens GLM-5.3 weights for coding and vulnerability hunting

Z.ai released the weights for its GLM-5.3 model on August 28, a 756 GB package with a broad commercial license, claiming steep gains in coding and vulnerability-hunting benchmarks. The model scored 28.3 on Terminal-Bench 3.0 (up from 4.6 for GLM-5.2), 66.9 on DeepSWE v1.1 (up from 46.2), and 54.4% on ExploitBench (up from 24.4%), with improvements attributed entirely to post-training. Z.ai founder Tang Jie had outlined autonomous agents as the next step for the GLM family in an April 2026 statement.

Z.ai opens GLM-5.3 weights for coding and vulnerability hunting The 756 GB release carries a broad commercial license and Z.ai's claims of steep gains on exploitation benchmarks. By Ryan Merket /author/ryan-merket · Published Primary source: Z.ai on X https://x.com/Zai org/status/2093354097122455713 Why it matters GLM-5.3 turns Z.ai's claimed advances in long-running coding and exploit-chain reasoning into downloadable infrastructure, with a license that leaves most deployment controls to users. Z.ai https://z.ai/blog/glm-5.3 released the weights for GLM-5.3 /models/z-ai/glm-5.3 on August 28th, turning its latest agentic coding and cybersecurity model into a package developers can download, run and modify rather than access solely through Z.ai's hosted products. The release extends the model-building strategy laid out by Z.ai founder Tang Jie. In an April 2026 founder's statement https://www1.hkexnews.hk/listedco/listconews/sehk/2026/0419/2026041900085.pdf , Tang described autonomous agents capable of long-running work as the next step for the GLM family. GLM-5.3 applies that thesis to two tasks where autonomy carries immediate value and risk: completing complex engineering projects and finding exploitable software flaws. Z.ai said in a post on X https://x.com/Zai org/status/2093354097122455713 that GLM-5.3 is its most capable model for agentic coding and cyber defense. The release itself is verifiable: the GLM-5.3 repository https://huggingface.co/zai-org/GLM-5.3 contains 756 GB of model files split across 141 Safetensors shards, alongside configuration files, deployment instructions and Z.ai's custom license. The configuration describes a mixture-of-experts model with 256 routed experts, eight experts selected per token and a maximum position setting of 1,048,576 tokens. Z.ai lists support for Transformers, vLLM, SGLang and other inference frameworks, giving developers several routes to self-host the model. The repository also provides an FP8-quantized release, although a 756 GB download still places full-scale deployment firmly in data-center territory. The gains came from post-training Z.ai says GLM-5.3 uses the same base model as GLM-5.2 /models/z-ai/glm-5.2:free , with the reported improvements coming entirely from post-training. That distinction matters because Z.ai is presenting reinforcement learning environments, task verification and long-horizon agent training as the source of the model's progress rather than another increase in base-model scale. On Z.ai's reported evaluations, GLM-5.3 scored 28.3 on Terminal-Bench 3.0, up from 4.6 for GLM-5.2, and 66.9 on DeepSWE v1.1, compared with 46.2 for the earlier model. Z.ai also reported a 50% improvement on its private Z.ai Code Bench, where GLM-5.3 completed more tasks while generating fewer output tokens. Those figures come from Z.ai's model card and technical report, and the disclosed evaluation settings vary by benchmark. Terminal-Bench 3.0, for example, was run with a 400,000-token context, as many as 600 agent turns and a 10-hour timeout per rollout. The resulting scores measure an agent operating inside a substantial harness, rather than a bare model answering short coding prompts. A broad license for a cyber model Cybersecurity is the sharper part of the release. Z.ai says it added vulnerability-discovery data and executable security environments during post-training, producing a model that became better at linking individual flaws into multi-stage exploitation plans. Z.ai reported an 84.5% score on CyberGym, up from GLM-5.2's 77.2%. On ExploitBench, which Z.ai describes as a deeper test of reasoning about real vulnerabilities and exploitation, GLM-5.3 reached 54.4%, compared with 24.4% for GLM-5.2. Z.ai's own table also shows closed models scoring substantially higher on that test, including 78% for Fable 5 and 76.5% for GPT-5.6 Sol /models/openai/gpt-5.6-sol:batch . Z.ai says work with security teams in China produced 2,436 vulnerability findings across 269 open-source projects, including 1,097 categorized as critical or high severity. Its technical blog listed 53 findings as publicly disclosed and 2,383 as still under embargo when the figures were published. Those totals remain Z.ai's claims, with most of the underlying findings unavailable for outside inspection while disclosure processes continue. The GLM-5.3 license https://huggingface.co/zai-org/GLM-5.3/blob/main/LICENSE broadly permits users to copy, modify, distribute, sell, deploy and fine-tune the weights. Its main commercial restriction applies to operators of model-as-a-service businesses whose corporate group generates more than $10 billion over a consecutive 12-month period; those operators must pass a Z.ai security review before commercial use. The clause does not impose the same review on smaller model hosts or on products that embed GLM-5.3 inside specific features. That structure gives independent developers and security researchers considerable freedom, while also placing most decisions about acceptable deployment on whoever downloads the model. With GLM-5.3, Z.ai is distributing the capability it spent its post-training budget sharpening. The weights release lets developers test the coding claims on their own infrastructure. It also makes the cyber capability available outside Z.ai's API controls.