GLM-5.3: Post-Training Only, and an Unexpected Cyber Result

What GLM-5.3 is

GLM-5.3 was released by Z.ai on 18 August 2026. It uses the same base model as GLM-5.2, and Z.ai are unusually direct that every gain comes from post-training — more environments, more diverse tasks, more compute spent on the training stack they had already built. That stack has three named parts: IndexShare for efficient long-context processing, SAO for reinforcement learning on long-horizon tasks, and slime for large-scale asynchronous training. It carries a 1 million token context window, 128,000 maximum output tokens, text-only input, and reasoning that is always on across three effort levels.

The coding claim, with the table behind it

Z.ai call GLM-5.3 the most capable open-weights model for coding, citing a 50% improvement over GLM-5.2 on their in-house Z.ai Code Bench and open-source state of the art on Terminal Bench 3.0 and Agents' Last Exam. They publish the comparison table rather than the headline alone, which makes it checkable. On Terminal Bench 2.1 it scores 88.2 against GLM-5.2's 81.0, with Kimi K3 at 88.3 and GPT-5.6 Sol at 88.8. Terminal Bench 3.0 is the more dramatic line: 28.3 against GLM-5.2's 4.6, a sixfold jump from post-training alone, though Claude Fable 5 reaches 33.7 and Sol 34.6. DeepSWE v1.1 goes from 46.2 to 66.9. It tops the table on AutomationBench v1.0.6 at 48.2, ahead of every model listed including Sol at 45.8, and GDPval-AA v2 at 1769, ahead of Fable 5 at 1743 and Sol at 1730.

The cyber result, and where it stops

Z.ai report that cyber capability developed faster than they expected as post-training scaled — their word is emergent — and that GLM-5.3 is state of the art on CyberGym for vulnerability discovery. The table supports that specific claim: 84.5 on CyberGym, the highest figure in it, above Fable 5 at 83.8 and Sol at 83.6. On ExploitGym it records 105 at two hours and 130 at six, which more than doubles GLM-5.2's 29 and 39 — but sits far behind Fable 5 at 181/247 and Sol at 216/293. On ExploitBench it scores 54.4 against Sol's 76.5.

Weights

Z.ai stated at launch that they would release the weights two weeks after, once safety evaluation and hardening were complete — which would put that around late August 2026. This site's roster still records the licence as not published and the model as not self-hostable, reflecting the position at launch, with GLM-5.2 noted as the version you can actually run yourself. Check Z.ai's Hugging Face presence for the current status. The same caution applies to pricing: the figures this site holds come from aggregated third-party listings rather than Z.ai's own page, and the API was documented as available soon rather than available.

Independent figures

On LMArena's text leaderboard GLM-5.3 has 1481 Elo (rank 28, max effort); by category, 1520 Elo in Coding (rank 36, max effort), 1499 Elo in Math (rank 20, max effort), 1521 Elo in Expert (rank 20, max effort) and 1457 Elo in Creative Writing (rank 32, max effort). SimpleQA Verified is low at 41. It generates 93.2 tokens per second with a 23.4 second time to first token, and roughly 41,107 output tokens across 420 seconds on a measured task.

Official Resources