Mystery solved: Ox-Alpha NiuLai large model is China's AI: Zhipu achieves computational power independence.

The hottest large model over the past week is undoubtedly the mysterious Ox-Alpha, an AI dubbed the "Ox Model" that has sparked widespread speculation. It has now been officially confirmed as a new GLM product from Chinese AI company Zhipu AI.

The specific name of this product has not yet been announced, and it could be GLM-5.3 Flash. Zhipu AI has confirmed it as their product, and the weights will be open-sourced tonight.

The overall performance of this Ox Model has not surpassed GLM-5.3, but it is multimodal, supporting images and video, and its overall performance remains highly competitive. It can be viewed as an enhanced version of DS V4 Flash.

Earlier speculation about the ownership of the Ox Model included guesses that it was Xiaomi's Mimo V3 Pro, Google's Gemini, and even some who analyzed it seriously and concluded it was Apple's AI model. Ultimately, it came back to Zhipu AI, which had already been identified through model fingerprinting.

Compared to the performance of the Ox Model, the most significant breakthrough this time is the computing power behind it. When Zhipu AI launched the Ox Model for free, they announced daily support of 100T of computing power. To put that in perspective, just one platform, OpenRouter, saw 17.5 trillion tokens of usage in seven days, higher than the V4 Flash that had been dominating the charts.

OpenCode's usage has also been staggering, reaching 5.1 trillion tokens in a single day on August 22. Other channels like Hermes and OMP also contributed, but even combined, these haven't yet reached the 100T usage Zhipu AI promised.

This implies that Zhipu AI now has abundant computing resources. Based on a daily call volume of 2.5 trillion tokens, the computing power required is equivalent to 2.3 EFLOPS, which translates to 1,300-2,100 B200 GPUs.

But Zhipu AI has promised 100T of daily computing power, which would require approximately 60,000-90,000 B200 GPUs. That is staggering. Considering Zhipu AI operates more than just one large model, their total computing power far exceeds what outsiders imagined.

Earlier reports indicated that Zhipu AI launched an AI data center with 1GW of computing power, using fully domestic chips. If it is indeed operational, then the free week of the Ox Model was just a small-scale test.

In summary, the Ox Model is not just another high-performing large model in China's AI field; the computing resources it represents are the core breakthrough. As long as computing power no longer lags too far behind companies like OpenAI and Anthropic, domestic AI large models will only catch up to frontier capabilities faster.