ByteDance founder says: No “distillation” of AI models According to reports, ByteDance will not resort to “distillation” technology as a shortcut to improve the capabilities of its AI models, even if it means the company is currently lagging behind, founder Zhang Yiming said at a staff meeting last month.

ByteDance founder says: No “distillation” of AI models According to reports, ByteDance will not resort to “distillation” technology as a shortcut to improve the capabilities of its AI models, even if it means the company is currently lagging behind, founder Zhang Yiming said at an employee meeting last month. This meeting is a plenary meeting of the ByteDance AI team (internally codenamed “Seed”). Zhang had rarely spoken at Seed team meetings before, but he broke his silence after other leadership discussed why the company wasn't using distillation to get ahead. Zhang Yiming said ByteDance "should be willing to sacrifice some short-term benefits for longer-term goals," according to employees, but he did not elaborate on what the long-term goals were. According to employees, ByteDance’s reluctance to use distillation technology is believed to be one of the reasons why its large-scale language models lag behind other competitors. By contrast, ByteDance is recognized as the global leader in AI video generation models, with its Seedance 2.0 model causing a stir earlier this year for its ability to generate ultra-realistic videos. ByteDance has increasingly become an outlier in China’s AI landscape, in part because the social media giant is the country’s only major AI company that uses almost exclusively proprietary models, while its peers focus on open source models, the report said. ByteDance’s flagship large-scale language model does not always appear on major global large-model rankings, which are a common window for outsiders to track model capabilities. Customers access Bytedance's model primarily through the company's cloud computing platform, which has a smaller market share in China than industry leaders Alibaba, Huawei and Tencent. This makes it difficult for non-customers to assess the actual performance of ByteDance's model. ByteDance’s latest large model, Seed 2.1 Pro, did not appear on Artificial Analysis, the most watched large model competition platform. The model ranks 19th on Arena AI’s web development coding rankings, well below other Chinese models. According to ByteDance employees, there has been a long-standing debate within the ByteDance Seed team about whether the company should move to distillation technology. This year, the rapid progress of domestic competitors’ open source models has put greater pressure on ByteDance. Every time a new powerful Chinese open-source model is released, ByteDance's internal debate over distillation heats up, according to employees. To be fair, American developers also affirmed the technological progress made by Chinese AI companies in improving models. For example, DeepSeek is widely recognized for its architectural innovation that significantly reduces memory requirements during model training and inference. Related articles: ByteDance founder Zhang Yiming makes his first public appearance in recent years via cnBeta.C