GLM-5.3 is Z.ai's 753B Mixture-of-Experts model for complex software engineering and long-horizon agentic work, with 40B active parameters. It is a text-only model released under the GLM-5.3 License.

Use GLM-5.3 in LM Studio Bionic

GLM-5.3 is available as a Cloud model in LM Studio Bionic, with up to 500K tokens of context in the current LM Studio Cloud configuration. The model architecture supports contexts up to 1M tokens.

Built for long-horizon agentic engineering

GLM-5.3 uses the same base model as GLM-5.2, with its gains coming from scaled post-training. Z.ai expanded training toward realistic units of expert work involving full codebases, documentation, testing tools, research environments, and multi-step workflows.

The model is designed to take ownership of sustained tasks end to end: understanding a codebase, planning changes, implementing them, running tests, and verifying results. It carries forward GLM-5.2's reinforcement-learning techniques, including compaction for long trajectories, while scaling the diversity and complexity of its training environments.

Reasoning effort

GLM-5.3 always operates with reasoning enabled. It supports three reasoning effort levels: low, high, and max. Z.ai recommends max for difficult coding and long-horizon work, while lower levels trade some computation for reduced latency and token use.

Security analysis

Z.ai reports that GLM-5.3's post-training gains extend to vulnerability discovery and multi-stage exploitation analysis. On CyberGym, it scores 84.5 compared with 77.2 for GLM-5.2. On ExploitBench, which evaluates reasoning about real vulnerabilities and their exploitation, it scores 54.4 compared with 24.4 for GLM-5.2.

These are security capabilities reported by Z.ai under the evaluation settings described in its launch post and model card. Model outputs still require expert review and responsible use.

Benchmark results

Z.ai reports broad improvements over GLM-5.2 across coding, agentic, and security evaluations:

Benchmark	GLM-5.3	GLM-5.2	Kimi K3	Claude Opus 4.8
Terminal-Bench 2.1	88.2	81.0	88.3	85.0
Terminal-Bench 3.0	28.3	4.6	17.4	21.1
DeepSWE v1.1	66.9	46.2	67.5	58.0
FrontierSWE	78.1	67.5	—	66.5
SWE-Marathon v1.1	42.5	19.4	48.1	48.8
PostTrainBench	39.8	31.7	32.0	32.9
CyberGym	84.5	77.2	80.0	78.1
ExploitBench	54.4	24.4	32.2	40.0
Toolathlon Verified	73.0	59.9	76.5	76.2
AutomationBench v1.0.6	48.2	26.2	46.7	41.0
Agents' Last Exam	28.5	23.8	27.6	25.7

Benchmark image and results are from Z.ai's GLM-5.3 launch post and official model card. Evaluation settings vary by benchmark; see the original sources for complete methodology and comparison details.

Sources and license

GLM-5.3 is released under the custom GLM-5.3 License. See Z.ai's launch post, model documentation, and official model card for model details, usage guidance, and evaluation methodology.