Model card briefing
GLM-5.3 uses the same base model as GLM-5.2—the publisher attributes every gain to post-training. Compared with GLM-5.2, it targets substantial improvements in complex coding and long-horizon tasks.
- Stronger coding: a reported 50% improvement over GLM-5.2 on the in-house Z.ai Code Bench, plus open-source SOTA claims on Terminal Bench 3.0 and Agents’ Last Exam.
- Emergent cyber capability: the README reports SOTA CyberGym vulnerability-discovery performance and more than double GLM-5.2’s results on exploitation benchmarks.
Extended README benchmark matrix
| Benchmark | GLM-5.3 | GLM-5.2 | Kimi K3 | DS-V4 Pro-0813 | Qwen3.8-Max | Opus 4.8 | Fable 5 | GPT-5.6 Sol |
|---|---|---|---|---|---|---|---|---|
| Terminal Bench 2.1 | 88.2 | 81.0 | 88.3 | 87.9 | 86.6 | 85.0 | 88.0 | 88.8 |
| Terminal Bench 3.0 | 28.3 | 4.6 | 17.4 | – | – | 21.1 | 33.7 | 34.6 |
| DeepSWE (v1.1) | 66.9 | 46.2 | 67.5 | 62.7 | 56.6 | 58.0 | 69.7 | 72.7 |
| NL2Repo | 58.0 | 48.9 | 58.0 | 61.1 | 55.9 | 69.7 | – | – |
| ProgramBench (Almost Solved) | 19.0 | 9.5 | 17.5 | – | 10.5 | 15.5 | 33.0 | 23.0 |
| FrontierSWE | 78.1 | 67.5 | – | – | – | 66.5 | 88.2 | – |
| SWE-Marathon (v1.1) | 42.5 | 19.4 | 48.1 | – | – | 48.8 | 33.1 | 42.5 |
| PostTrainBench | 39.8 | 31.7 | 32.0 | – | – | 32.9 | 41.8 | 36.2 |
| CyberGym | 84.5 | 77.2 | 80.0 | 83.3 | 78.5 | 78.1 | 83.8 | 83.6 |
| ExploitGym (2h / 6h) | 105 / 130 | 29 / 39 | 36 / 70 | – | 14 / 26 | 80 / 120 | 181 / 247 | 216 / 293 |
| ExploitBench | 54.4 | 24.4 | 32.2 | – | 28.8 | 40.0 | 78.0 | 76.5 |
| Toolathlon Verified | 73.0 | 59.9 | 76.5 | 74.1 | 72.5 | 76.2 | 74.7 | 74.9 |
| AutomationBench (v1.0.6) | 48.2 | 26.2 | 46.7 | 43.2 | 39.8 | 41.0 | 46.2 | 45.8 |
| Agents’ Last Exam (ALE-CLI) | 28.5 | 23.8 | 27.6 | 25.7 | 27.0 | 25.7 | 23.8 | 28.6 |
| HLE w/ Tools | 62.5 | 54.7 | 59.8 | 60.0 | 56.2 | 57.9 | 63.9 | 64.5 |
| GDPval-AA v2 | 1769 | 1508 | 1682 | 1590 | 1739 | 1588 | 1743 | 1730 |
Serve GLM-5.3 locally
The extended card provides deployment guidance for SGLang, vLLM, TokenSpeed, Transformers, KTransformers, and Unsloth. It also notes vLLM-Ascend, xLLM, and SGLang support for Ascend NPU deployments.
Reasoning and chat-template controls
- reasoning_effort: accepts low, high, and max. It defaults to max when omitted or passed another value; keep max for benchmark reproduction.
- clear_thinking: defaults to false in the chat template. The README explicitly recommends clear_thinking=true for chat scenarios.
Extended evaluation protocols
General reasoning, repository, and software benchmarks
HLE w/ tools: temperature 1.0, top_p 0.95, maximum 163,840 generated tokens and 300K managed context, with GPT-5.6-luna (medium) as judge. NL2Repo: temperature 1.0, top_p 1.0, 64K new tokens and 1M context, with rule- and LLM-based anti-hacking checks. DeepSWE: mini-swe-agent, temperature 0.95, top_p 1.0, six-hour timeout and 400K context. FrontierSWE: Proximal evaluation with 1M context, max effort and 128K output. PostTrainBench: Claude Code 2.1.207, max effort, temperature 1.0, top_p 1.0, 128K output and 1M context, weighted over three runs. SWE-Marathon: max effort, temperature 1.0, top_p 0.95, 128K output and 1M context.
Terminal, agent, and automation benchmarks
Terminal-Bench 2.1: Claude Code 2.1.207, temperature 1.0, top_p 1.0, 65,536 new tokens and six-hour timeout. Terminal-Bench 3.0: max reasoning, 400K context, 128K output, avg@3, isolated official task containers, 600-turn cap and ten-hour timeout. Agents’ Last Exam: official protocol with Claude Code, max reasoning, 1M context and 64K output across 105 isolated tasks. Toolathlon Verified: official service, pass@1 averaged over three runs. AutomationBench: v1.0.6 including the null-type handling fix. GDPval-AA v2: evaluated by Artificial Analysis.
Cybersecurity benchmark controls
CyberGym: Claude Code 2.1.207, max reasoning, no web tools, temperature 1.0, top_p 1.0 and 128K output; single-run Pass@1 over 1,507 tasks with domain allowlisting. ExploitGym: 869 tasks under two- and six-hour inference-time budgets, single-run Pass@1, no web tools and domain allowlisting. ExploitBench: 300 interaction-round limit, with average coverage over 41 tasks across three revisions and domain allowlisting.
Technical report: GLM-5 Team, “GLM-5: from Vibe Coding to Agentic Engineering,” 2026 · arXiv:2602.15763.