Zhipu (02513.HK) said it released GLM-5.3 today, using the same base model as
GLM-5.2 and attributing all performance gains to post‑training optimization. The
company said GLM-5.3 improves performance on complex coding and long‑horizon
tasks, delivering a 50% uplift versus GLM-5.2 on its internal Z.ai code
benchmark. Zhipu called GLM-5.3 the most capable open‑source weight model and
said network‑capability gains during scaled post‑training deployment exceeded
expectations. On the CyberGym platform, GLM-5.3 led vulnerability discovery
tests, with improvements concentrated in later stages of exploit chains and more
than double the exploitation‑benchmark performance of GLM-5.2. Pending security
evaluation and hardening, Zhipu plans to publish the model weights two weeks
after release.
