Zhipu AI Launches GLM-5.3 with 743B Parameters, Secretly Tested on OpenRouter as Ox Alpha

Researchers revealed that Chinese company Zhipu AI secretly tested its new model, GLM-5.3, on the OpenRouter platform under the alias Ox Alpha. The model's identity was confirmed with 99% confidence through response fingerprinting analysis.

The model boasts 743 billion parameters and supports a context window of up to one million tokens, achieving a 50% improvement over its predecessor in a performance leap that positions it competitively alongside the world's largest language models.

This covert testing approach reflects a growing trend among AI companies, which resort to open platforms to gather real-world performance data without prematurely disclosing the model's identity.

The model's size, margin of improvement, and secret testing methodology place Zhipu AI in a strong position, as these combined indicators suggest that GLM-5.3 is gearing up for serious competition in the large language model market upon its official release.

What do these terms mean?

Model Fingerprinting: A technique relying on distinct patterns in a model's responses to identify it, even when using an alias.

1M Token Context: The model's ability to process up to one million text units at a single time, equivalent to roughly 750,000 words or several long novels.

Shadow Testing: Deploying a model with a hidden identity in a real-world environment to collect actual performance data without an official announcement.