ByteDance founder Zhang Yiming made an unusual call in 2023: he banned his AI researchers from distilling rival models, a technique that has become central to the US-China AI dispute. That early decision, made years before Washington threatened sanctions over model theft, now sets ByteDance apart as the only major Chinese AI lab that Anthropic has not accused of extracting proprietary data from competitors.

The timing matters because it suggests either remarkable foresight or a genuine research philosophy. Distillation involves feeding queries to a powerful model like Claude, collecting the outputs, and using them to train a cheaper system that mimics the original. It is a shortcut that becomes especially tempting when export controls limit access to computing power, as they do for Chinese firms.

Why Is ByteDance's Restraint So Unusual?

In July 2026, US Treasury Secretary Scott Bessent warned that Chinese firms could face financial sanctions and blacklisting over what he characterized as model theft. The White House had already reclassified covert distillation by adversarial states as a national security threat in April. Yet ByteDance's policy predates all of this by three years.

Anthropic, the maker of Claude, accused three major Chinese labs of coordinated extraction campaigns. DeepSeek, Moonshot AI, and MiniMax allegedly used roughly 24,000 fraudulent accounts to generate more than 16 million exchanges with Claude. Alibaba's Qwen operation faced even larger accusations, with some 28.8 million exchanges using nearly 25,000 fraudulent accounts. ByteDance alone escaped public accusation.

The irony is sharp: ByteDance's restraint has not translated into market dominance. Its best models have been surpassed by smaller Chinese labs. Moonshot released Kimi K3, a 2.8-trillion-parameter open-weight model that developers ranked above GPT-5.6 Sol and Fable 5 on a blind coding leaderboard, despite Moonshot being a fraction of ByteDance's size.

What Does ByteDance's Commercial Position Look Like Today?

On the user adoption front, ByteDance's Doubao reached 345 million monthly active users by March 2026, well ahead of Qwen at 166 million and DeepSeek at 127 million. Globally, Doubao ranked second only to ChatGPT by app users. This suggests that abstaining from distillation has not crippled ByteDance's ability to build a user base, even if its models lag on technical benchmarks.

The company faces unique exposure to US regulatory action. TikTok, ByteDance's flagship product, has spent years under American political pressure. Any accusation of intellectual property theft could compound that vulnerability, making the 2023 distillation ban a form of insurance.

Beijing is pressing its labs to build better models with less compute, a demand shaped by export controls rather than preference. Under those constraints, distillation is an obvious shortcut. Refusing it is costly.

How Are Chinese Open-Weight Models Reshaping the Competitive Landscape?

• MiniMax's Video Breakthrough: MiniMax released H3, a multimodal model capable of generating text, images, video, and audio, on July 31, 2026. The company's stock surged 78.21% in the week following the launch, with H3 achieving an Elo score of 1,242 for text-to-video with audio and ranking first globally for video editing.
• Open-Source Ecosystem Impact: When MiniMax open-sourced H3 on August 3, 16 chip manufacturers including Huawei Ascend, AMD, and Intel completed integration simultaneously, along with development communities like Hugging Face and ModelScope. Over 100 companies launched on day one, mirroring the ecosystem expansion that DeepSeek triggered in early 2025.
• Token Consumption and Revenue Growth: Video generation models are expected to drive significant token consumption. A single 15-second video can consume approximately 310,000 tokens in real-world testing, compared to 30 to 100 tokens for a typical chat. This high-density token use case mirrors the revenue growth pattern that made Anthropic's coding tools so valuable.

The pricing advantage is also notable. H3 costs less than one-third the price of flagship video models at 2K resolution and less than half the price of mainstream models at 768P resolution.

Capital markets are beginning to price in video generation as a new center for token consumption, similar to how coding tasks validated the commercialization path for text and code work. If video generation moves from experimentation into production workflows for advertising, e-commerce, short-form dramas, and gaming, it could become a major revenue driver.

What Does This Mean for the Broader US-China AI Competition?

The strategic question now is whether Chinese labs can sustain their pace of improvement without borrowing from American systems. ByteDance has effectively volunteered to answer that question through abstention. Whether the 2023 distillation ban proves to be principle, foresight, or an expensive mistake depends on results that remain uncertain.

The Economist argued this month that US models still offer better value but that convergence is likely to continue. If that holds, the competitive advantage will shift toward whichever ecosystem can build better models faster with constrained resources.

Meanwhile, Beijing is tightening its own grip on model access. China has weighed curbing overseas access to its top AI models and issued guidance in July requiring export licenses for model weights and training pipelines leaving the country. This suggests that the open-weight movement in China may face new restrictions even as companies like MiniMax push toward greater ecosystem openness.