Z.ai GLM-5.3 reaches 84.5% on CyberGym vulnerability benchmark LLM performance via agentic fine-tuning, surpassing GLM-5.2 and proprietary rivals while triggering safety delays.

GLM-5.3 tokens surpass one billion in personal use, with AveryChing highlighting strong coding throughput and value on the new open-weight release from Z.ai.

Z.ai GLM-5.3-Flash activates 18B of 320B params via MoE, runs on domestic silicon under export curbs, MIT license, 1M context.

GLM-5.2 now live on B.AI API with 1M context window for complex reasoning and agent workflows from Z.ai.