Z.ai Launches GLM-5.3 With Gains in Coding and Cybersecurity

Z.ai has released GLM-5.3, the latest version of its open-weights model. The company said GLM-5.3 uses the same base model as GLM-5.2, with all of its performance gains coming from post-training.

GLM-5.3 is available through Z.ai's Coding Plan and ZCode, as well as coding agents including Claude Code and OpenCode. Z.ai said the model weights will be publicly released two weeks after launch, following safety evaluation and hardening.

Z.ai said GLM-5.3 delivers a 50% improvement over GLM-5.2 on its in-house Z.ai Code Bench and achieves what the company describes as open-source state-of-the-art results on public benchmarks including Terminal-Bench 3.0 and Agents’ Last Exam. 

On Terminal-Bench 3.0, GLM-5.3 scored 28.3, compared with 4.6 for GLM-5.2, while its DeepSWE v1.1 score increased to 66.9 from 46.2.

The company said the gains came from scaling the post-training system used for GLM-5.2. Z.ai continued scaling IndexShare for long-context processing, SAO for reinforcement learning on long-horizon tasks and slime for large-scale asynchronous training, while adding more environments, more diverse tasks, and more compute.

Z.ai said cybersecurity capabilities developed faster than expected as post-training scaled. The company said GLM-5.3 moved beyond identifying individual vulnerabilities and began reasoning across multiple stages of an exploitation chain.

On CyberGym, GLM-5.3 scored 84.5%, up from 77.2% for GLM-5.2. Z.ai said this was the best result on the benchmark. On ExploitBench, GLM-5.3 scored 54.4%, more than twice GLM-5.2's 24.4% score.

On ExploitGym, GLM-5.3 completed 105 tasks under the two-hour budget and 130 under the six-hour budget, compared with 29 and 39 for GLM-5.2. Z.ai said these budgets are normalised using each model's inference speed, so the figures are not simple real-time comparisons.

Z.ai also tested GLM-5.3 against real-world codebases with several security teams in China. After expert review, screening, and deduplication, the company said the model identified 2,436 vulnerabilities across 269 open-source projects, including 1,097 medium-to-high-severity issues.

Z.ai’s Security Disclosure Ledger breaks those 1,097 findings down into 107 critical and 990 high-severity vulnerabilities. The findings covered system kernels, operating systems, browser engines, open-source infrastructure, web applications, and network protocols, with the oldest flaw dating back to 1981.

Z.ai said the work has become an ongoing disclosure program through its Security Disclosure Ledger, which tracks affected projects, severity ratings and CVE identifiers where available. At the time of the announcement, the ledger listed 53 publicly disclosed findings and 2,383 under embargo.

Z.ai said the model weights will be publicly released two weeks after launch, following safety evaluation and hardening.