Z.ai

Z.ai, formerly Zhipu AI, published GLM-5.3 on 14 August 2026 in an unsigned blog post with no technical report and no model card. Same base model as GLM-5.2: the lab attributes the entire gain to post-training, "Scaling post-training is all we did." Terminal-Bench 3.0 moves from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9, and Z.ai Code Bench, an in-house private benchmark, gains 50 percent over GLM-5.2 while cutting output tokens. The claim of the most capable open-weights coding model does not follow from the lab's own table: against Kimi K3 the score is 3-3 with one tie.

Working with équipes de sécurité chinoises non nommées, the model found 2,436 vulnerabilities in 269 open source projects, the oldest introduced in 1981, average lifetime 26.6 years. The Z.ai Security Disclosure Ledger lists 53 disclosed and 2,383 under embargo, a 2.2 percent publication rate. Z.ai presents the cyber capability as emergent while stating that vulnerability-discovery data and environments were deliberately added to the training mix. Weights were promised within two weeks, once safety evaluation and hardening finish: a date, with no definition of hardening, no third-party evaluator, and no condition that would withhold publication.

Artificial Analysis ranked GLM-5.2 first among open-weights models and third overall on GDPval-AA at 1524 Elo. Anthropic's September 2026 threat intelligence report named Z.ai among seven Chinese labs practising distillation, cleaning reasoning traces from Opus 4.8 across 273 accounts and 3.4 million exchanges, and abandoning Fable for models judged less protected. Almost every GLM-5.3 evaluation was run inside Claude Code 2.1.207.