ByteDance’s Founder Rules Out Distillation on AI Models
The Takeaway
- ByteDance founder Zhang Yiming rejects AI model distillation strategy.
- Decision aims to avoid U.S. scrutiny and protect TikTok’s global operations.
- ByteDance’s LLMs currently lag rivals due to this strategic choice.
ByteDance won’t resort to distillation as a shortcut to advancing its AI model capabilities, even if that means the company lags behind domestic rivals for now, founder Zhang Yiming told employees during a meeting last month. One reason, say people close to the company, is ByteDance’s complicated history with the U.S. government over the company’s TikTok video app.
The previously unreported comments, heard by two employees present at the meeting, offer a rare glimpse into the ByteDance founder’s thinking on the company’s AI development strategy. The meeting, an all-hands for ByteDance’s AI team internally known as Seed, took place just after Moonshot AI’s Kimi K3 model sparked alarms in the U.S. that Chinese models were narrowing the gap at a pace quicker than previously expected.
In the weeks since, a senior White House official has accused Kimi K3 of achieving better performance by distilling U.S. frontier models. (Distillation is the process of training a new model using outputs from existing, most advanced models.)
The caution against distillation stems from the company’s experiences with TikTok, which the U.S. government threatened to ban multiple times over national security concerns, according to one of the employees and another person with knowledge of ByteDance’s internal discussions.
TikTok’s existential crisis was resolved last year after a deal was reached to sell the vast majority of TikTok’s U.S. data security operation to a group of U.S.-led investors. But ByteDance continues to run the commercial operations of TikTok globally, including the U.S. And the company worries that distilling U.S. frontier models would set off a new wave of scrutiny from Washington and pose new threats to TikTok, according to the employee and the person with knowledge of ByteDance’s internal discussions.
Zhang, who rarely spoke at the Seed team’s previous meetings, broke his silence after other leaders talked about why the company won’t use distillation to get ahead. ByteDance “should be willing to sacrifice some short-term gains for longer-term goals,” Zhang said, without elaborating what the long-term goals are, according to the employees.
ByteDance’s reluctance to use distillation has been seen as a reason for why its large language models are trailing other AI labs in China, including Alibaba and upstarts like DeepSeek, Z.ai and Moonshot, according to the employees. In contrast, ByteDance is widely recognized as a global leader in AI video generation models, following the success of its Seedance 2.0 model, which made shockwaves earlier this year for its ability to produce hyperrealistic videos.
ByteDance has increasingly become an oddity in China’s AI landscape, partly because the social media giant is the only major Chinese AI firm whose LLMs are almost all proprietary models, while its peers focus on open-source models.
ByteDance’s flagship LLMs are not always included in major global LLM leaderboards, a common window for outsiders to track model capabilities. Customers access ByteDance’s models mainly 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 gauge how good ByteDance’s models are.
ByteDance’s latest LLM, Seed 2.1 Pro, is not seen on Artificial Analysis, the most widely followed leaderboard. The model ranks No. 19 on Arena AI’s coding leaderboard for web development, far below other Chinese models.
Inside ByteDance’s Seed team, there have long been ongoing debates about whether the company should turn to distillation, according to the ByteDance employees.
This year, rapid advancements of open-source models from domestic rivals have created more pressure for ByteDance. Every time a new powerful Chinese open-source model dropped, ByteDance’s internal debates about distillation heated up, according to the employees.
Distillation has become a sensitive and highly controversial topic due to the intensifying AI rivalry between the U.S. and China. Trump administration officials have long criticized Chinese AI developers for allegedly using distillation to copy the capabilities of leading models from Anthropic and OpenAI.
Anthropic has accused many Chinese AI firms—DeepSeek, Moonshot, MiniMax, Z.ai and Alibaba—of illicitly extracting Claude’s capabilities through large-scale distillation to improve their own models. Anthropic, however, hasn’t made such an accusation against ByteDance. The accused Chinese firms haven’t publicly commented about the distillation allegations.
To be sure, developers in the U.S. have also credited the Chinese AI firms with making their own technical advances to improve their models. Distillation alone cannot explain the strong performance of China’s best open-source models such as Moonshot’s Kimi K3, which rivals the best U.S. models on some benchmarks, they said. DeepSeek, for example, is widely recognized for its architectural innovations that drastically lower memory requirements during model training and inference.
But researchers at Chinese AI firms also acknowledge that distillation does bring significant advantages. It saves a lot of training time and computation overhead because a new model under development can learn directly from the reasoning steps of an existing frontier model, instead of relying solely on raw and curated training data, they say.
