ByteDance founder Zhang Yiming has said the company will not rely on AI distillation to advance its models, even if that means falling behind competitors. The remarks, first reported by The Information and carried by the American Bazaar on 5 August 2026, put one of China's largest technology firms on the record at a moment when distillation sits at the centre of the dispute between Washington and Chinese AI laboratories.
What Zhang told staff
People familiar with the matter said the decision is partly influenced by ByteDance's complicated relationship with the US government over its ownership of TikTok. Zhang, who has kept a low profile in recent years, is described as becoming more hands-on with the company's AI work, meeting senior researchers and monitoring recruitment. The South China Morning Post has reported that he has been sitting in on the Seed team's core technology reviews and reading OpenAI papers late into the night.
Distillation, in this context, means training a model on the outputs of a more advanced model, typically one built by another company. At small scale the technique is a routine and openly practised part of the industry and is generally considered legal. The current controversy concerns covert, industrial-scale use of a rival's outputs.
The dispute surrounding the pledge
Anthropic has accused several Chinese AI firms, including DeepSeek, Moonshot, MiniMax, Z.ai and Alibaba, of distilling its Claude models. ByteDance has not been accused. The sharpest confrontation concerns Moonshot AI's Kimi K3, a 2.8-trillion-parameter open-weight model aimed at software engineering and autonomous task execution. Michael Kratsios, director of the White House Office of Science and Technology Policy, alleged in an X post that Moonshot distilled Anthropic's Fable model through a purpose-built internal platform that rotated access routes to evade detection, using NVIDIA GB300 servers in Thailand that are barred from export to China under US rules. One report states that Anthropic traced 3.4 million Claude conversations to Moonshot through fake accounts. Moonshot has not publicly responded to the allegations.
Washington has paired such accusations with policy threats. Treasury Secretary Scott Bessent has floated sanctions should Chinese companies adopt distillation on an industrial scale, and a bipartisan House bill would make it easier to identify and sanction distillation attempts. One report, citing Stanford's Institute for Human-Centered AI, states that leading US and Chinese models are now broadly comparable despite US AI spending outstripping China's by a reported 23 to one.
Why ByteDance can afford the stance
ByteDance's position is underwritten by Seedance 2.0, the audio and video model it rolled out on its CapCut editing platform under the Dreamina brand earlier in 2026. According to one account, the company backed scaling the model past 200 billion parameters despite internal disagreement over cost, and the product has since produced gross margins of roughly 70 to 90 per cent while driving more than half of Volcano Engine's model-as-a-service revenue.
The competitive pressure is nonetheless real. Alibaba has unveiled its largest model to date, Qwen3.8-Max, sending its shares higher, and a research firm estimated that DeepSeek's latest model is priced more than 100 times cheaper than Anthropic's Claude Fable 5.
The verification problem
Analysts cited in the reporting note that a no-distillation pledge is difficult to verify from outside the company. Neither Zhang nor ByteDance has explained how the policy will be enforced internally, whether it covers synthetic data generated by ByteDance's own smaller models, or which performance benchmarks Zhang is prepared to sacrifice in the near term. The pledge is therefore a statement of intent rather than an auditable commitment.
What is established and what is merely claimed
Established: Zhang Yiming has said ByteDance will not rely on distillation, according to The Information's reporting. ByteDance is not among the firms Anthropic has accused of distillation. US officials, including Kratsios and Bessent, have publicly raised distillation as a sanctions and legislative issue, and a bipartisan House bill exists to that end.
Merely claimed: the allegation that Moonshot distilled Anthropic's Fable model to build Kimi K3, including the Thailand servers and access rotation, remains an accusation by a White House official, and Moonshot has not responded. The reported tracing of 3.4 million Claude conversations rests on a single outlet's account. Whether ByteDance's pledge can be enforced or verified is unknown, and the assessment that top US and Chinese models are now broadly comparable is an outside estimate, not a settled fact.
Sources
- American Bazaar report on Zhang Yiming's remarks, 5 August 2026 · American Bazaar
- Cryptopolitan report on the pledge, Seedance and the wider dispute · Cryptopolitan