GLM-5.3 Model Card (Z.ai, August 2026)

The GLM-5.3 model card is the README of the zai-org/GLM-5.3 repository on Hugging Face, published by Z.ai when it released the weights of GLM-5.3 under the GLM-5.3 License; the repository's first public commit is dated August 27–28, 2026, two weeks after the model's API launch on August 14 (Source: huggingface.co). Like the GLM-5.2 card it is a model card rather than a system card: it reports capability benchmarks, including three offensive-security benchmarks, but no refusal, misuse-safeguard or alignment evaluations.

Model description

The card states that GLM-5.3 uses the same base model as GLM-5.2 and that "every gain comes from post-training." It makes two claims. On coding, Z.ai calls GLM-5.3 the most capable open-weights model, reports a 50% improvement over GLM-5.2 on its in-house Z.ai Code Bench without giving absolute scores, and claims the best open-source results on Terminal Bench 3.0 and Agents' Last Exam. On security, under the heading "Emergent Cyber Capability," it says cyber capability "developed faster than we expected" as post-training was scaled, that GLM-5.3 leads on CyberGym for vulnerability discovery, and that it more than doubles GLM-5.2 on exploitation benchmarks. The card notes that the reasoning_effort parameter accepts low, high and max, defaults to max, and should be left at the default when reproducing benchmark results.

Evaluation approach

The benchmark table has eight columns — GLM-5.3, GLM-5.2, Kimi K3, DeepSeek-V4 Pro-0813, Qwen3.8-Max, Opus 4.8, "Fable 5 (w/ fallback)" and GPT-5.6 Sol — and marks the best figure in each row in bold. GLM-5.3 is bolded on three rows: CyberGym, AutomationBench and GDPval-AA v2.

• Coding and long-horizon engineering: Terminal Bench 2.1 and 3.0, DeepSWE v1.1, NL2Repo, ProgramBench (reported as the share of tasks "Almost Solved"), FrontierSWE, SWE-Marathon v1.1 and PostTrainBench. Seven of the footnotes name Claude Code 2.1.207 as the harness, generally at maximum reasoning effort. Terminal-Bench 3.0 is reported as the average of three rollouts per task with a 600-turn cap and a 10-hour timeout. FrontierSWE was run by Proximal, with a Dominance score as of August 14, 2026. For PostTrainBench and SWE-Marathon the footnotes disclose changes to the benchmarks' own checks: pattern-matching checks for third-party API use and an import-detection anti-cheat check were replaced with inspection by a language model, because Z.ai found they produced false positives.
• Cybersecurity: CyberGym (single-run pass@1 over 1,507 tasks with no per-task time limit), ExploitGym (tasks solved out of 869 under two-hour and six-hour budgets) and ExploitBench (average capability coverage over 41 tasks, taking the union of capabilities reached across three revisions, with at most 300 interaction rounds). Each was run in Claude Code without web tools and with a domain whitelist that the card says is meant "to prevent the agent from cheating"; for CyberGym, Git-related information was also removed and the agent was placed inside the task container. The ExploitGym budgets are not wall-clock limits: Z.ai computed them from API inference time rescaled by each model's tokens-per-second rate as published by Artificial Analysis, and states that it ran GLM-5.3, Kimi K3 and Qwen3.8-Max this way.
• Agentic and knowledge work: Toolathlon Verified (official service, pass@1 averaged over three runs), AutomationBench v1.0.6, Agents' Last Exam in its command-line form (105 tasks, official evaluators), Humanity's Last Exam with tools (300,000-token context with a context-management strategy, GPT-5.6 Luna as judge) and GDPval-AA v2, which the card says was evaluated by Artificial Analysis.

Several GLM-5.2 figures differ from those in the GLM-5.2 card because the benchmark version or metric changed, for example SWE-Marathon (13.0, then 19.4 on v1.1) and ProgramBench (63.7, then 9.5 "Almost Solved"). The card does not identify the source of the comparison models' figures beyond the ExploitGym footnote.

Reported results

Evaluation results the document states for which there is no separate benchmark page; results on benchmarks that have a page are recorded there and cite this page. Scores and settings are as printed in the document.

Provenance

The raw file is the verbatim README fetched from Hugging Face on September 19, 2026; the repository was last modified on September 4, 2026. A benchmark chart embedded as an image is not reproduced. The results are developer-reported or relayed by the developer, which is the basis for the medium confidence rating. The same table appears in Z.ai's launch post (Source: z.ai). The companion card for GLM-5.3-Flash gives evaluation footnotes but presents its results only as an image (Source: huggingface.co).

Relationships

• supports: GLM-5.3 — primary reference for the model's reported benchmark results.
• depends-on: GLM-5.3 License — the licence the weights were released under; GLM-5: from Vibe Coding to Agentic Engineering (Zhipu AI, 2026) — the technical report the card cites.
• related: GLM-5.2 Model Card (Z.ai, June 2026), CyberGym, ExploitGym, ExploitBench, Zhipu AI