Z.ai

Z.ai, formerly Zhipu AI, publishes the «GLM» family of open-weights models, and three releases from that line surface here within four months. «GLM-5.1» (754B MoE, 40B active, MIT) led SWE-Bench Pro at 58.4% in April 2026, alongside an API price rise of roughly 40%. «GLM-5.2», released 16 June 2026 at 753B parameters with 40B active and a 1M-token context, scored 1524 Elo on GDPval-AA, third overall behind Claude Fable 5 and Claude Opus 4.8, and well clear of the next open model, MiniMax-M3, at 1408. Mistral AI then chose «GLM-5.2» as the first third-party open model on its regional endpoints.

The «GLM-5.3» announcement of 14 August 2026 arrived as an unsigned blog post with no technical report and no model card. Its methodological claim is narrow and checkable: the same base model as «GLM-5.2», with the entire gain from post-training. "Scaling post-training is all we did." Terminal-Bench 3.0 moves from 4.6 to 28.3, and the headline benchmark, «Z.ai Code Bench», is in-house and private. The "most capable open-weights model for coding" claim is not settled by Z.ai's own table: against Kimi K3 the score is 3-3 with one tie.

The security numbers are what make this entity hard to file. Working with unnamed Chinese security teams, Z.ai reports 2,436 vulnerabilities found across 269 open source projects, the oldest introduced in 1981, mean lifetime 26.6 years. The «Z.ai Security Disclosure Ledger» shows 53 disclosed and 2,383 under embargo. Weights were promised within two weeks, once safety evaluation and hardening finish. That is a date, not a criterion.