GLM-5.3

2026-08-14 · Publisher : Z.ai · Domain : Artificial Intelligence

Z.ai launched GLM-5.3 on August 14, 2026, same base as GLM-5.2 but with strengthened post-training: better long-form coding, cybersecurity (CyberGym 84.5). Open weights on August 25, under its own GLM-5.3 license, not MIT.

### Context Two months after GLM-5.2, the vendor reuses the same base and concentrates its effort on post-training. The closed models compared are Claude Opus 4.8, "Fable 5 (with fallback)" and GPT-5.6 Sol, and the open ones Kimi K3, DeepSeek-V4 Pro-0813 and Qwen3.8-Max. The post specifies that the weights come out two weeks after launch, once the safety evaluation and hardening are complete. ### What the model brings - A 50% improvement over GLM-5.2 on Z.ai Code Bench (internal benchmark) and an open state of the art on Terminal Bench 3.0 (28.3 vs 4.6 for GLM-5.2) and Agents' Last Exam, according to the vendor. - A cyber capability described as emergent: state of the art on CyberGym (84.5) for vulnerability discovery, with stronger gains in exploitation (ExploitGym at 2 h: 105 vs 29 for GLM-5.2). - The post publishes a register: 2,436 vulnerabilities tracked, 53 publicly disclosed, 2,383 under embargo, including 1,097 critical or high, across 269 open source projects; the release notes speak of parity with "Mythos 5" in code review and vulnerability discovery (vendor claim, not verified). - The reasoning_effort parameter (low, high, max; max by default) to control the thinking budget. ### Architecture and specifications - Same base as GLM-5.2 (MoE glm_moe_dsa, 256 experts of which 8 active, 78 layers; about 753.3 billion parameters by the HF count). Context: 1M tokens; maximum output not read for this model; modality: text. - LICENSE: "GLM-5.3 License", specific to the vendor. It allows use, copying, modification, distribution and sale, but requires a security review by Z.ai before any commercial use for a licensee (or its affiliates) operating a "Model as a Service" and exceeding USD 10 billion in cumulative revenue over 12 months. It is therefore no longer the MIT license of versions 5 to 5.2. - Weights: zai-org/GLM-5.3 (and BF16), published on August 25, 2026 on Hugging Face. ### Cost and access | Model (Z.ai API, current pricing) | Input /M | Cached input /M | Output /M | |---|---|---|---| | GLM-5.3 | USD 1.4 | USD 0.26 | USD 4.4 | ### Benchmarks | Benchmark | GLM-5.3 | GLM-5.2 | Claude Opus 4.8 | Fable 5 (with fallback) | GPT-5.6 Sol | |---|---|---|---|---|---| | Terminal Bench 2.1 | 88.2 | 81.0 | 85.0 | 88.0 | 88.8 | | Terminal Bench 3.0 | 28.3 | 4.6 | 21.1 | 33.7 | 34.6 | | DeepSWE (v1.1) | 66.9 | 46.2 | 58.0 | 69.7 | 72.7 | | CyberGym | 84.5 | 77.2 | 78.1 | 83.8 | 83.6 | | ExploitGym (2 h / 6 h) | 105 / 130 | 29 / 39 | 80 / 120 | 181 / 247 | 216 / 293 | | Toolathlon Verified | 73.0 | 59.9 | 76.2 | 74.7 | 74.9 | | HLE with tools | 62.5 | 54.7 | 57.9 | 63.9 | 64.5 | | GDPval-AA v2 | 1769 | 1508 | 1588 | 1743 | 1730 | Source: Hugging Face model card and z.ai post for GLM-5.3, scores published by the vendor; GDPval-AA v2 is measured by Artificial Analysis, FrontierSWE by Proximal. ### Versus closed models On Terminal Bench 2.1, GLM-5.3 (88.2) outperforms Claude Opus 4.8 (85.0) and Fable 5 with fallback (88.0), and stays 0.6 points behind GPT-5.6 Sol (88.8). On CyberGym, it leads the row (84.5 vs 83.8 and 83.6). In exploitation, the gap with the two best closed models is clear: 105 vs 181 and 216 at two hours on ExploitGym. On Terminal Bench 3.0, it remains 5.4 points behind Fable 5 (28.3 vs 33.7) and 6.3 points behind GPT-5.6 Sol (34.6). The API is priced at USD 1.4 input and USD 4.4 output per million tokens; no contemporary closed pricing has been verified. For the open or closed choice, the license change matters: commercial use is free for the vast majority of players, but there is a prior security check for the very large providers of models as a service. ### Key takeaways GLM-5.3 brings an open model closer to the closed frontier level in coding and cybersecurity, without a new base. It is also the first version of the 5 series outside the MIT license, with a USD 10 billion revenue threshold for "Model as a Service" providers.

Tags : LLM, Open Source, Z.ai, Zhipu AI, GLM, MoE, Cybersécurité, Licence GLM-5.3