Chinese AI lab's open-weight model trails the best closed model by just 1% on key benchmarks, and Vercel integrated it within three days of release
Guillermo Rauch, the CEO of Vercel and creator of Next.js, said he was “genuinely impressed, almost shocked” by the coding abilities of GLM-5.2, a new open-weight AI model built by Chinese lab Z.ai. That kind of reaction from someone who runs one of the most widely used developer platforms in the world tends to make people pay attention.
The model dropped on June 13, 2026. By June 16, Vercel had already integrated it into its AI Gateway. A three-day turnaround from release to production integration is not normal.
What makes GLM-5.2 different #
On FrontierSWE, one of the more demanding evaluations for coding AI, GLM-5.2 trails Claude Opus 4.8, the leading closed model, by just 1%.
The model scored 81.0 on Terminal-Bench 2.1, a massive jump from its predecessor GLM-5.1, which managed only 63.5. That’s roughly a 28% improvement in a single generation. Internal benchmarks paint an even more dramatic picture, with app-dev task scores leaping from 21 out of 70 to 48 out of 70.
GLM-5.2 supports 1 million tokens, up from 200K in GLM-5.1. A five-fold increase means it can handle sprawling codebases and long-horizon engineering tasks that would have overwhelmed its predecessor.
Z.ai made the model available through Hugging Face and local GGUF formats, which means developers can run it on their own hardware without needing API access or subscriptions.
The open-source AI arms race heats up #
GLM-5.2 leads all open-source models in long-horizon coding evaluations. Jeremy Howard, a prominent figure in the AI community, has highlighted the model’s capacity to rival closed systems.
Vercel’s rapid integration tells its own story. Rauch’s company serves as infrastructure for hundreds of thousands of developers. Adding GLM-5.2 to the AI Gateway within 72 hours of release isn’t just an endorsement. It’s a bet that this model will see meaningful production usage.
What this means for developers and investors #
If an open-weight model can perform within 1% of the best closed alternative on serious coding benchmarks, the reasons to pay premium API prices for closed models start looking thinner. Especially when you can run the open model locally, inspect its weights, and fine-tune it for your specific use case. GLM-5.2 represents maybe the most aggressive single-generation leap in that direction for coding-specific tasks, with Terminal-Bench scores jumping from 63.5 to 81.0 and app-dev task scores more than doubling from 21 to 48 out of 70.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our