Summary
Fahd Mirza delivers a hands-on test of GLM 5.2, the latest model from Chinese lab ZAI (Zhipu AI), released the same evening Anthropic’s Fable 5 and Mythos 5 were pulled from global access by US export controls. The video frames this launch as part of a deliberate pattern: Chinese AI labs rapidly filling the vacuum created by US regulatory restrictions on frontier models.
GLM 5.2 runs on a 744-billion-parameter mixture-of-experts architecture and supports a true 1-million-token context window — five times larger than its predecessor. It offers two thinking effort levels (high and max), is text-only at launch, and will be released under an MIT open-source license within a week of recording. Access is currently available via the ZAI API. Mirza integrates the model with his Hermes agent framework on Ubuntu for testing.
The benchmark is deliberately non-trivial: a real SSL certificate monitoring application spanning more than 4,000 files across three directories. In a single prompt, GLM 5.2 is asked to read the entire codebase, identify and fix bugs, and add a brand-new desktop notification feature that fires alerts when certificates approach expiration. The model completes the full task in under seven minutes — writing new files, wiring them into the existing architecture, running its own tests, and verifying the notification feature end-to-end in the browser. Mirza positions GLM 5.2 alongside Kimi K2.7 as one of the strongest open-weight options for long-horizon agentic coding tasks.
📺 Source: Fahd Mirza · Published June 14, 2026
🏷️ Format: Hands On Build
