# ByteDance's Zhang Yiming Bans AI Model Distillation Even as Rivals Race Ahead > Source: > Published: 2026-08-10 07:20:24+00:00 *ByteDance is trying to make a virtue out of restraint: Zhang Yiming wants Seed researchers to avoid distilling rival models just as Washington is turning that practice into a sanctions issue.* ByteDance founder Zhang Yiming has put a hard line around how the company's Seed AI team trains models. According to the Financial Times, ByteDance is developing a giant model while refusing to use distillation from existing systems, even though that choice may slow it down against Chinese rivals such as Moonshot AI and DeepSeek. That's the point. Distillation is fast. It lets a smaller model learn from the outputs of a stronger one without paying the full cost of building every capability from scratch, which is why the argument over it has moved from research labs into trade policy. The timing is awkward for a reason. Reuters reported on July 22 that Michael Kratsios, director of the White House Office of Science and Technology Policy, accused Moonshot AI of using Anthropic's Fable model to help build Kimi K3. He said the US government had information that Moonshot distilled Fable for K3, and he also raised questions about access to Nvidia GB300 systems. Treasury Secretary Scott Bessent then posted on X: "Open source is not open season on American IP," warning that sanctions and Entity List designations could follow when firms run covert, industrial-scale distillation attacks. The risk is legal. ## Zhang Is Choosing The Slower Route Zhang's public posture looks less like a research preference and more like a business hedge. The Information has reported that ByteDance's internal caution around distillation dates back to 2023, so this isn't a rule invented after Washington named Moonshot. ByteDance is simply making the rule more visible when the cost of looking careless has gone up. If you're ByteDance, that matters more than it would for a smaller lab. TikTok's US business is still tied up in national security politics, and a company that has spent years trying to prove it can be trusted by American regulators has no reason to invite a fresh fight over American model output. That is a hedge. The Financial Times reported that ByteDance's model could reach as many as 10 trillion parameters, larger than Anthropic's Mythos estimate of about 8 trillion and far above Moonshot's Kimi K3. The project is still in pre-training, a phase the FT said usually lasts three to six months, so nobody should treat it as a finished answer to US frontier models yet. Seed is already a serious operation, and ByteDance's Doubao assistant has hundreds of millions of monthly users in China. This isn't retreat. It's a decision to spend more time and compute now in return for a cleaner story later. ## The Moonshot Fight Changed The Stakes Anthropic laid the groundwork for this fight in February, when it said it had found industrial-scale campaigns by DeepSeek, Moonshot and MiniMax to extract Claude capabilities. The company said the three labs generated more than 16 million exchanges through about 24,000 fraudulent accounts, breaching its terms of service and regional access rules. Moonshot's share, according to figures reported from Anthropic's disclosure, was about 3.4 million exchanges focused on agentic reasoning, tool use, coding, data analysis, computer-use agents and computer vision. ByteDance was not named in that disclosure. That absence helps. But the Fable-to-K3 claim is still a hard claim, not a proved public case. TechCrunch reported that Fable had been publicly available only since July 1 before Moonshot released K3 later that month, which leaves a narrow window for the model to have been mainly built by copying Fable. Some researchers have questioned whether distillation alone could explain K3's capabilities in that period. A model this large doesn't appear from a few clever prompts. Washington may have private evidence, but the public case still rests on official accusation rather than published logs or a technical report. No sanction has landed. The Information reported that discussions about adding Moonshot to the Commerce Department's Entity List were preliminary, while the Bureau of Industry and Security was investigating whether Chinese firms had access to advanced US chips. That is still enough to change behavior. For ByteDance, the punishment doesn't have to arrive before the calculation changes: the threat alone can make a faster training shortcut look expensive. It is a trade. Frankly, Zhang's line puts pressure back on the rest of China's AI race. If ByteDance, with its cash, users and engineering bench, is willing to move slower to avoid the appearance of copying US models, rivals can't hide forever behind the idea that everyone does it. Distillation is a real method, and used on your own models it can be perfectly normal. The better move is to say exactly where the boundary is and stay well clear of it. It may cost ByteDance a few leaderboard points this summer, but it gives the company something more useful if Washington keeps pushing: a cleaner answer when regulators ask how the model was built. **Also read:** [TSMC's July sales jumped 45% as AI chip demand keeps outrunning tariffs](https://startupfortune.com/tsmcs-july-sales-jumped-45-as-ai-chip-demand-keeps-outrunning-tariffs/) • [DeepSeek Warns Its Ultra-Cheap V4-Flash Pricing Won't Last](https://startupfortune.com/deepseek-warns-its-ultra-cheap-v4-flash-pricing-wont-last/) • [An AI Assistant Booking a Gym Class in Melbourne Ended Up Hacking the Site](https://startupfortune.com/an-ai-assistant-booking-a-gym-class-in-melbourne-ended-up-hacking-the-site/)