Z.ai released GLM-5.3, a post-training-only upgrade over GLM-5.2 that shares the same base model but delivers large gains on coding and long-horizon agentic benchmarks (e.g., Terminal-Bench 3.0 up from 4.6 to 28.3, DeepSWE v1.1 up from 46.2 to 66.9). The release highlights emergent cybersecurity capabilities: on ExploitBench the model's score more than doubled to 54.4%, and in real-world testing with security teams it found 2,436 vulnerabilities across 269 projects, some dating back roughly 40 years, now tracked in a public Z.ai Security Disclosure Ledger. The model runs on the open-source slime RL framework with Megatron and SGLang, achieving over 2.3x higher training throughput. Notably, disabling the 'thinking' mode is no longer supported — applications using thinking.type: disabled must migrate to enabled with reasoning_effort: low before switching to glm-5.3, or requests will fail. Weights will be open-sourced within two weeks.
Table of contents
Stronger CodingEmergent Cyber Capabilityslime : Built for Long-Horizon RL ScalingGetting started with GLM-5.3FootnotesQuestions this post answers
Do I need to change my API config when upgrading from GLM-5.2 to GLM-5.3?
Yes, if the application currently sets thinking.type to disabled, it must be changed to enabled and reasoning_effort set to low before switching the model ID to glm-5.3, otherwise the request will fail. GLM-5.3 no longer supports disabling the thinking mode; the model exposes three effort levels instead: low, high, and max, with max recommended for coding tasks. Track breaking model migration requirements like this one on daily.dev before they break your integration.
How much better is GLM-5.3 than GLM-5.2 at finding software vulnerabilities?
GLM-5.3 scores 84.5% on CyberGym versus 77.2% for GLM-5.2, and more than doubles GLM-5.2's ExploitBench score, reaching 54.4% versus 24.4%. On ExploitGym it completes 105 exploitation tasks within two hours and 130 within six hours, up from 29 and 39 for GLM-5.2, though it still trails closed models like Mythos 5 at 181 and 247 tasks. Developers evaluating AI security tooling follow model capability jumps like this on daily.dev.
How many real-world vulnerabilities has the GLM model found in production codebases?
Working with security teams on real-world codebases, the model identified 2,436 vulnerabilities across 269 projects, including 1,097 medium-to-high severity issues, spanning system kernels, operating systems, browser engines, open-source infrastructure, web applications, and network protocols. The oldest flaw dated back roughly 40 years, with vulnerabilities living an average of 26.6 years before discovery. Security teams weighing AI-assisted vulnerability discovery track findings like these on daily.dev.
33 Impressions1 Comment