ByteDance dumps distillation, plans 5-trillion-parameter model

ByteDance is making aggressive moves to reshape its AI strategy. Founder Zhang Yiming appeared at a July all-hands meeting for the Seed division and declared that ByteDance will not use distillation — training on outputs from advanced models like Claude or GPT — as a shortcut, even if that means its models temporarily lag behind domestic competitors. CEO Liang Rubo reinforced this at an August 5th all-hands meeting, emphasizing independent R&D and long-term optimization. To support this direction, ByteDance has reportedly established a new first-level 'AI Data and Security' department, integrating a data team of more than 1,000 people and giving it equal standing with Seed, Flow, and Douyin. This places data governance and security at the core of its AI operations rather than an afterthought.

Meanwhile, ByteDance is reportedly discussing training a model with over 5 trillion parameters, which would make it the largest known model in China by parameter scale. If realized, this signals a massive compute investment and a move to compete on foundational research rather than quick iteration. In parallel, Doubao's commercialization is accelerating: it launched three monthly subscription tiers (68, 200, and 500 yuan), began charging commissions on life-service orders, and merged the Feishu product team with Doubao to focus on desktop Agent products. Collectively, ByteDance appears to be reorganizing its AI business from architecture to monetization, sacrificing short-term benchmark wins for a slower but potentially more defensible position in China's AI race.