The Powerful Stealth AI Model ‘Ox Alpha’ Is Now GLM-5.3-Flash, and You Can Use It

The Chinese company Z.ai has taken a once anonymous AI model called Ox Alpha and renamed it GLM-5.3-Flash. This move solidifies Z.ai's position in the AI landscape, as the new model is both powerful and cost-effective, rivaling popular models from Anthropic and OpenAI. Ox Alpha quickly gained popularity on OpenRouter, with all its traffic served through Chinese-made AI chips. Z.ai's stock price rose when it was revealed to be the creator of the model.

Ox Alpha was introduced as a "frontier model built for efficient coding, sustained agentic work, and real-world production use," according to an OpenRouter announcement. Z.ai confirmed that the model wouldn't be trained on user prompts, providing an added layer of security for users. Z.ai's GLM family consists of open-weight models, meaning that core components are publicly released, allowing users to download and utilize them independently.

In contrast, closed-weight models, like those from big US developers such as OpenAI and Anthropic, keep all components, including source code, weights, and training data, under lock and key.

GLM-5.3-Flash's weights are accessible on Hugging Face, another unified AI platform. The open-weight model stands in contrast to closed-weight models from major US companies, which has sparked heated debates in the US due to the potential misuse of such models. However, major American companies, including OpenAI, have expressed support for open ecosystems. Meta recently announced its intention to release the weights of a previously closed model, Muse Spark 1.2.

GLM-5.3-Flash boasts enhanced agentic capabilities, enabling the model to utilize a browser and a computer to navigate web pages and interact with desktop applications. Z.ai provides access to the model through its GLM coding plans, starting at $18 per month. The model also offers three times the usable quota of GLM-5.3. Since GLM-5.3-Flash is open-weight, users can download and use it on their own hardware, though doing so may necessitate substantial computing power, as the model weights themselves are over 300GB in size.