Zhipu AI Releases GLM-5.3-Flash Model Following Viral Stealth Trial
Beijing-based Zhipu AI has officially released its GLM-5.3-Flash model, which was previously tested under the code name 'Ox Alpha.' The company claims the model successfully processed 62 trillion tokens on a cluster of 100,000 domestic Chinese chips.
Zhipu AI, a prominent Chinese artificial intelligence startup, has officially launched its latest open-weight model, GLM-5.3-Flash. The release follows a period of intense speculation surrounding a mysterious model known as 'Ox Alpha,' which gained significant traction on AI marketplaces like OpenRouter and OpenCode. According to the company, Ox Alpha was a stealth trial for the new GLM-5.3-Flash system.
Zhipu AI reported that the model processed 62 trillion tokens during its week-long trial period. A central claim of the announcement is the model's reliance on domestic infrastructure; the company stated that the entire system operated on a cluster of 100,000 Chinese-produced chips. This development is being viewed as a significant milestone for China's domestic semiconductor and AI industries, which have been seeking to reduce reliance on foreign hardware, particularly high-end GPUs from companies like Nvidia.
The announcement has led to a positive market reaction, with Zhipu AI shares seeing an increase following the news. While the company has provided data regarding the model's performance and the scale of the chip cluster used, independent verification of the internal hardware performance remains limited to the company's own disclosures. The transition from the viral 'Ox Alpha' moniker to the official GLM-5.3-Flash branding marks the end of a high-profile testing phase that drew significant attention from the global AI development community.
📡 Media Analysis
How each outlet framed the story — angles, word choices, and what they chose to push or ignore.
Focused on the technical achievement and the strategic importance of domestic chip usage.
"high-profile stealth trial"
✓ Only outlet to report: The specific detail that the model processed 62 trillion tokens during the trial period.
🔍 What Nobody's Reporting
- ·Lack of independent verification regarding the performance of the domestic chips compared to international equivalents.
- ·No details provided on the specific architecture or manufacturer of the 100,000 domestic chips.
📰 Sources
0 A-rated source(s) among 1 total. Lowest trust: SCMP (B)