ByteDance, the world’s most valuable private technology company, is navigating a rare moment of public vulnerability in its artificial intelligence ambitions. At an all-hands meeting on August 6, CEO Liang Rubo acknowledged that the gap between ByteDance’s Seed large language model and the leading overseas systems is widening. He asked the team to accept being behind for a period of time and to commit to independent research and development. The admission, reported by 36kr and confirmed by multiple Chinese technology media outlets, stands in sharp contrast to the company’s usual posture of confident forward momentum.
The disclosure arrived alongside a separate revelation from LatePost that ByteDance is discussing training a large language model with a parameter scale exceeding 5 trillion, a figure that would dwarf the 2.4 trillion parameters of Alibaba’s Qwen3.8-Max and the 2.8 trillion parameters of Moonshot’s Kimi K3. The two signals are not contradictory but complementary: Liang is resetting the near-term catching-up timeline, while the 5 trillion parameter plan defines the method. ByteDance is not prepared to close the gap by borrowing from competitors’ outputs. Instead, it is concentrating computing power, data, and organizational resources on a larger-scale pre-training run that it hopes will leapfrog the current frontier.
The No-Distillation Principle
The strategic choice not to use distillation is the most distinctive and consequential aspect of ByteDance’s AI development philosophy. Distillation, the practice of feeding queries to a powerful model, collecting the outputs, and using them to train a cheaper system that approximates the original, has become the central flashpoint in the US-China AI dispute. Anthropic has accused DeepSeek, Moonshot AI, and MiniMax of using roughly 24,000 fraudulent accounts to generate more than 16 million exchanges with Claude in a coordinated extraction campaign. Alibaba’s Qwen operation was accused of an even larger effort: 28.8 million exchanges using nearly 25,000 fraudulent accounts.
ByteDance is conspicuously absent from Anthropic’s list of accused parties. According to The Next Web, Zhang Yiming reportedly forbade the company’s AI researchers from distilling rival models as far back as April 2023, years before avoiding U.S. retaliation became a commercial priority. The policy has been enforced through API call inspection and output similarity sampling. Since then, Seed has internally debated the question multiple times, and the answer has always been no.
(Related: ByteDance’s Doubao Powers Nubia’s Navix Ultra as AI Agent Smartphones Multiply at WAIC)
What Abstaining Has Cost
The no-distillation principle has come at a real commercial price. ByteDance’s best language models have been surpassed by smaller Chinese rivals that were not bound by the same constraints. Its Doubao-Seed-Code model and Trae coding tool have underperformed against Zhipu’s GLM 5 and Moonshot’s Kimi K2 in the coding benchmarks that matter most to enterprise customers. Zhipu’s annual recurring revenue has exceeded $1 billion; Moonshot’s has exceeded $300 million. ByteDance’s AI revenue is structurally skewed toward multimodal generation. Seedance and Seedream together account for more than half of Doubao’s token consumption, rather than the higher-margin language model and coding services where competitors have been winning.
The financial stakes are significant. Volcano Engine, ByteDance’s cloud and AI services arm, is targeting 400 billion yuan in revenue this year, up from approximately 150 billion yuan in 2025. Token consumption for the Doubao large model reached 180 trillion in June, below the internal target of 250 to 300 trillion. As the short drama industry that fueled Seedance’s growth begins to plateau, the revenue gap created by the weakness in language models is becoming harder to ignore.
The 5 Trillion Parameter Bet
The response is a high-stakes wager on scale. The proposed 5 trillion parameter model will be led by Xiang Liang, head of Seed Foundation, and Shen Ke, head of large language model pre-training data, both veterans of ByteDance’s search, advertising, and recommendation systems. Zhang Yiming personally invited Guo Daya, a core researcher from DeepSeek, to join Seed and take charge of special training for coding capabilities. The company is also dismantling its traditional horse-racing organizational model, where multiple teams run parallel experiments in the same direction, in favor of a concentrated, single-route approach that allocates more resources to each bet but accepts less redundancy.
The analogy Zhang Yiming is drawing on is ByteDance’s own Seedance video model. In 2025, when most teams were shifting focus from pre-training to post-training for video generation, Seedance pushed pre-training to the extreme, building the first video generation model with a full Mixture-of-Experts architecture at 200 billion parameters.
After its launch in February 2026, it became the world’s highest-performing video model and the revenue engine of Volcano Engine’s model-as-a-service business. Zhang Yiming said ByteDance “should be willing to sacrifice part of short-term revenue for long-term goals.” Whether the 5 trillion parameter language model can replicate Seedance’s success, in a field where OpenAI, Anthropic, and xAI are all advancing simultaneously, is the question that will define ByteDance’s AI ambitions for the rest of the decade.
(Related: DeepSeek Signals a Significant Price Hike, Reversing Its Low-Cost Edge)