Z.ai GLM-5.3 Open-Weight Frontier Coding Model
Z.ai released GLM-5.3 on 14 August 2026, saying it uses the same base model as GLM-5.2 with every gain coming from post-training, that it is the most capable open-weights model for coding, and that its cyber capability grew faster than expected.
Tool Overview
Features, steps and FAQ below
Features
- ✓ Z.ai says GLM-5.3 uses the same base model as GLM-5.2, with every gain coming from post-training on its IndexShare, SAO and slime stack by scaling environments, tasks and compute
- ✓ Z.ai says it is the most capable open-weights model for coding, with a 50% improvement over GLM-5.2 on its in-house Z.ai Code Bench and open-source state of the art on public benchmarks including Terminal Bench 3.0 (28.3 vs 4.6) and Agents' Last Exam (28.5 vs 23.8)
- ✓ Z.ai says it improves markedly on complex coding: DeepSWE v1.1 rises from 46.2 to 66.9, and at Max effort it reaches 34.5% at roughly 75K output tokens per task, versus 23.4% at 96K for GLM-5.2, improving both performance and token efficiency
- ✓ Z.ai says its cyber capability developed faster than expected as training scaled: GLM-5.3 scores 84.5% on CyberGym (the best result, ahead of Mythos 5's 83.8%), 54.4% on ExploitBench (more than doubling GLM-5.2's 24.4%), and completes 105 tasks in two hours and 130 in six hours on ExploitGym (versus 29 and 39 for GLM-5.2)
- ✓ Z.ai says that on real-world codebases with security teams it identified 2,436 vulnerabilities across 269 projects (including 1,097 medium-to-high severity) and built the Z.ai Security Disclosure Ledger to track disclosure; it says the model weights will be released about two weeks after launch, once safety evaluation and hardening are complete
How to Use
- Call GLM-5.3 at Z.ai via the API (see the official docs at docs.z.ai/guides/llm/glm-5.3) or subscribe to the Z.ai Coding Plan
- Adopt GLM-5.3 for long-horizon coding and agentic tasks in real engineering environments such as CLIs and terminals, replacing older versions that need stronger coding capability
- Watch for and download the model weights opened about two weeks after launch (once safety evaluation and hardening are complete, per Z.ai)
- Check the Z.ai Security Disclosure Ledger to follow the vulnerabilities found and disclosed using the model
FAQ
What is GLM-5.3?
Z.ai says GLM-5.3 is its flagship model released on 14 August 2026, using the same base model as GLM-5.2 with every capability gain coming from post-training, and says it is the most capable open-weights model for coding.
What does GLM-5.3 improve over GLM-5.2?
Z.ai says coding performance improves 50% over GLM-5.2 on its in-house Z.ai Code Bench, with Terminal Bench 3.0 rising from 4.6 to 28.3, Agents' Last Exam from 23.8 to 28.5 and DeepSWE v1.1 from 46.2 to 66.9, while consuming fewer output tokens at the same performance.
Is GLM-5.3 open source?
Z.ai says the weights will be released for download about two weeks after launch, once safety evaluation and hardening are complete, and it cites the rapid growth of cyber capability as one reason for the delayed weight release.
How does GLM-5.3 perform on cybersecurity?
Z.ai says it scores 84.5% on CyberGym (the best result), 54.4% on ExploitBench (more than doubling GLM-5.2's 24.4%) and identified 2,436 vulnerabilities across 269 projects on real-world codebases.
How do I get started with GLM-5.3?
Z.ai says you can call it via the Z.ai API (see docs.z.ai/guides/llm/glm-5.3) and it is available to GLM Coding Plan users, or you can wait for the model weights opened about two weeks after launch to self-host.