Key facts
- Z.ai released GLM-5.3, a 743-billion-parameter model, on Thursday.
- The lab claims GLM-5.3 is the most capable open-weights model for coding.
- GLM-5.3 is built by scaling post-training on the GLM-5.2 base.
- The model demonstrates improved performance and token efficiency over GLM-5.2.
- GLM-5.3 leads in cybersecurity benchmarks, flagging 2,436 vulnerabilities.
- API access and downloadable weights will be released following safety reviews.
Chinese AI lab Z.ai has launched GLM-5.3, a 743-billion-parameter model that it is marketing as the leading open-weights solution for coding tasks. The model is accessible through the GLM Coding Plan and ZCode, with API access and downloadable weights to follow a safety review.
Z.ai stated that GLM-5.3 was developed through scaling post-training on the GLM-5.2 base, incorporating more diverse tasks and environments. The focus was on token efficiency, resulting in a model that consumes fewer tokens per task than its predecessor. GLM-5.3 achieved a 34.5% score on Z.ai's in-house Code Bench at maximum effort, using approximately 75,000 output tokens, compared to GLM-5.2's 23.4% score using 96,000 tokens.
While GLM-5.3 outperforms its predecessor and some open-weight competitors, it still trails behind closed U.S. models like Claude Opus 4.8 and GPT-5.6 Sol on certain benchmarks. On the Terminal Bench 3.0, GLM-5.3 scored 28.3, slightly behind Fable 5 (33.7) and GPT-5.6 Sol (34.6). On the DeepSWE v1.1 benchmark for fixing GitHub issues, it scored 66.9, behind Kimi K3 (67.5) and Fable 5 (69.7).
In cybersecurity, GLM-5.3 showed significant improvement, leading CyberGym at 84.5% and flagging 2,436 vulnerabilities across 269 open-source projects. The model's pricing is also a key differentiator, with Z.ai's API rates reportedly about a tenth of U.S. frontier models per token.
Z.ai, based in Beijing and on the U.S. Entity List, aims to provide a cost-effective alternative to U.S. models. The open-weights release is anticipated in approximately two weeks.