Headline
Zhipu released GLM-5.3 on August 14 with a single architectural change: none. The base model is the same ~743B-parameter stack as GLM-5.2. Every improvement is the product of post-training scaling on a new open-source framework called Slime.
Benchmarks vs GLM-5.2
| Benchmark | GLM-5.2 | GLM-5.3 | Delta | |---|---|---|---| | Terminal Bench 3.0 | 4.6 | 28.3 | 6x | | DeepSWE v1.1 | 46.2 | 66.9 | +44.8% | | Agents' Last Exam | 23.8 | 28.5 | +19.7% | | CyberGym | 77.2 | 84.5 | +7.3 | | AutomationBench | 26.2 | 48.2 | +22.0 | | HLE w/Tools | 54.7 | 62.5 | +7.8 | | GDPval-AA v2 | 1508 | 1769 | +261 |
GLM-5.3 ranks open-source #1 on six of seven agentic benchmarks and #2 on one. Terminal Bench 3.0, which measures whether a model can autonomously complete complex tasks inside a real terminal environment, is the headline jump: GLM-5.2 essentially could not operate a terminal, while GLM-5.3 approaches a full engineer's day of work.
Coding Efficiency
On Z.ai Code Bench, GLM-5.3 High completes a task at 31.4% accuracy using 50K tokens, versus Claude Opus 4.8 Max at 29.5% using 120K tokens. That is a 2.4x token-efficiency advantage with higher accuracy.
Because API pricing is token-priced, this matters commercially: Zhipu positions the 50K-vs-120K contrast as a direct competitive strike at Opus 4.8 pricing.
Post-Training Recipe
The Slime framework is open-source and introduces two scheduling components:
- IndexShare
- SAO (Stabilized Asynchronous Optimization, inferred from context)
- 2,404 potential vulnerabilities
- 1,088 medium/high-severity
- Oldest flaw: a 40-year-old architectural defect in DNS itself, where a malformed request can amplify 8,000,000x
- Zhipu official announcement
- Singularity.kiwi vulnerability analysis
- GitHub AI Daily 8-15
- Qiqi Chao hands-on review
- Same base, new ceiling: All GLM-5.3 gains come from post-training scaling on the unchanged 743B base; the Slime framework is itself open-sourced.
- Agentic benchmark leadership: 6 of 7 benchmarks at open-source #1; Terminal Bench 3.0 jumped 6x.
- Token efficiency beats Opus: 50K vs 120K tokens for comparable or higher accuracy on Z.ai Code Bench.
- Emergent security capability: 2,404 vulnerabilities including a 40-year-old DNS architecture flaw; coordinated-disclosure ledger at cvd.z.ai.
- Delayed weight release: Two-week hardening window before open-source, a first among Chinese frontier vendors.
These are used to scale reinforcement-learning training over long-horizon task environments for an unusually long duration. No base-model retraining occurred.
Emergent Cybersecurity Capability
Two weeks of pre-release red-team testing across 269 repositories surfaced:
CyberGym (white-box review) hit 84.5%, beating Mythos 5 (83.8%) and GPT-5.6 Sol (83.6%). ExploitBench scored 54.4, more than 2x GLM-5.2's 24.4.
Zhipu launched cvd.z.ai as a public coordinated-disclosure ledger tracking CVE status, severity, and vulnerability age. 2,383 of the 2,404 findings remain in coordinated disclosure.
Release Cadence
GLM-5.3 is live now on ZCode, AutoClaw, GLM Coding Plan, and developer platforms Trae, Qoder, JoyCode, OpenCode, CodeBuddy, and WorkBuddy. Full weights open-source in two weeks—a deliberate hardening window for safety teams, and the first time a Chinese frontier-model vendor has formalized release-then-harden-then-open-source as a standard pipeline.
Watch Items
1. Number of fine-tuned derivatives within 24 hours of weight release. 2. GLM Coding Plan subscription growth against Claude Code's ~70% weekly retention benchmark.