Same base model, more post-training: GLM-5.3 pushes coding and cybersecurity further

GLM-5.3 strengthens agentic coding and cybersecurity through post-training, with open weights planned after two weeks of evaluations.

GLM-5.3 uses the same base model as GLM-5.2. Z.ai attributes all of the gains to a larger post-training phase, with more environments, longer-running tasks, and more compute dedicated to reinforcement learning. The work notably builds on IndexShare for long-context processing, SAO for agentic trajectories, and slime for large-scale asynchronous training.

Coding is one of the main areas of improvement. GLM-5.3 rises from 4.6 to 28.3 on Terminal-Bench 3.0, from 46.2 to 66.9 on DeepSWE v1.1, and from 23.8 to 28.5 on Agents' Last Exam. On its internal Z.ai Code Bench, the company also reports higher success rates while using fewer output tokens than GLM-5.2. At maximum effort, the model reaches 34.5% with roughly 75,000 output tokens per task, compared with 23.4% and 96,000 tokens for its predecessor.

The RL environments are now designed to more closely resemble real-world units of work. Some tasks can include codebases, documentation, storage systems, compute clusters, and experimental results, with multiple stages of diagnosis, modification, testing, and verification. Z.ai also generates some of these environments and their verifiers, although the company notes that the process still requires a meaningful amount of human involvement.

The most significant gains appear in cybersecurity. After adding data and environments focused on vulnerability discovery, Z.ai says it observed faster-than-expected progress across exploitation tasks. GLM-5.3 reaches 84.5% on CyberGym, up from 77.2% for GLM-5.2, and 54.4% on ExploitBench, compared with 24.4%. On ExploitGym, it completes 105 tasks within a normalized two-hour budget and 130 within six hours, versus 29 and 39 respectively for GLM-5.2. Closed models, however, still maintain a substantial lead on the more advanced exploitation benchmarks.

Z.ai also says it tested its models with several security teams on real-world codebases. After review, filtering, and deduplication, the company reports 2,436 vulnerabilities across 269 projects, including 53 that had been publicly disclosed at the time of publication and 2,383 still under embargo. The findings are being progressively recorded in a Security Disclosure Ledger with information such as the affected project, severity, and CVE when available.

GLM-5.3 operates with reasoning permanently enabled and offers three effort levels: low, high, and max. The max setting is recommended for coding. The model is currently available through GLM Coding Plan and ZCode, including in environments such as Claude Code and OpenCode. The open weights are expected to be released two weeks after launch, once safety evaluations and hardening are complete.