Three days. That’s how long it took Z.AI’s newest model to climb near the top of one of the most competitive AI coding benchmarks around. GLM-5.2 (Max) hit the #2 spot on Arena.ai’s Code Arena: Frontend leaderboard by June 16, scoring between 1,593 and 1,595 points, trailing only Fable 5.

For context, it beat Claude Opus 4.7 (Thinking) by 29 points.

What GLM-5.2 actually is

Z.AI, the Chinese AI company formerly known as Zhipu AI, released GLM-5.2 (Max) around June 13. The model uses a Mixture-of-Experts (MoE) architecture, which is essentially a way of building very large models that only activate a fraction of their total capacity for any given task.

The total parameter count lands somewhere between 744B and 753B, but only about 40B parameters are active at any given time. That keeps inference costs manageable while still tapping into a massive knowledge base.

The model also ships with a 1M-token context window, optimized for long-horizon tasks including project-scale engineering and multi-step reasoning. GLM-5.2 is distributed under an MIT license, meaning anyone can run it locally, self-host it, fine-tune it, or build commercial products on top of it without licensing fees.

The leaderboard picture

Code Arena: Frontend isn’t the only benchmark where GLM-5.2 is making noise. The model claims the #1 position on Design Arena and ranks highly in Agent Arena. GLM-5.2 also shows strong results relative to Kimi-K2.6 and Minimax-M3, two other models that have been competitive in frontend coding tasks.

Why this matters beyond benchmarks

For developers, an MIT-licensed model can be deployed on-premise, integrated into proprietary workflows, and customized without worrying about API rate limits, usage-based pricing, or vendor lock-in. The MoE architecture delivers this while keeping active parameters at roughly 40B against a total parameter pool north of 744B, reducing compute costs for organizations that want to run models locally.