GLM

«GLM» is the model family published by «Z.ai», the Chinese lab formerly known as Zhipu AI, under open weights. Three dated releases hang off it in the source notes: «GLM-5.1» (April 2026), «GLM-5.2» (16 June 2026) and «GLM-5.3» (14 August 2026). The parent entity carries the lineage; the numbers belong to each release.

What makes the family worth following is the size of the claim it keeps testing: that open weights can sit next to proprietary frontier models on work that has economic value. Artificial Analysis measured «GLM-5.2» at 1524 Elo on GDPval-AA, third overall behind Claude Fable 5 (1783) and Claude Opus 4.8 (1615), and 116 points ahead of the next open model, MiniMax-M3. It is a 753-billion-parameter Mixture of Experts (40 billion active), 1M-token context, MIT licence, priced at $1.40 / $4.40 per million tokens.

«GLM-5.1» had already taken SWE-Bench Pro at 58.4%, while lagging on reasoning (GPQA Diamond 86.2% against Gemini 3.1 at 94.3%). «GLM-5.3» reused «GLM-5.2»'s base model entirely: the gain came from post-training, and Terminal-Bench 3.0 moved from 4.6 to 28.3.

The unresolved part is what «GLM-5.3» surfaced alongside the coding scores. Working with Chinese security teams, the model identified 2,436 vulnerabilities across 269 open source projects, of which 53 are disclosed and 2,383 remain under embargo. Z.ai promised weights "in two weeks, once safety evaluation and hardening are complete", naming a date rather than a criterion.