Skip to content

ByteDance Escalates AI Race With 10-Trillion Model

ByteDance is currently pre-training an artificial intelligence model featuring up to 10 trillion parameters. The initiative, first reported by the Financial Times, signals a major escalation in the global AI race, placing the TikTok parent company in direct competition with the largest frontier systems under development in the United States.

By

2 min read

bytedance-10-trillion-ai-model-concept
A stylized representation of the scale and complexity of the ByteDance 10-trillion parameter AI model.

ByteDance is currently pre-training an artificial intelligence model featuring up to 10 trillion parameters. The initiative, first reported by the Financial Times, signals a major escalation in the global AI race, placing the TikTok parent company in direct competition with the largest frontier systems under development in the United States.

At 10 trillion parameters, ByteDance’s new system marks a massive leap in hardware scale. Parameters represent the internal configuration variables an AI uses to process information, identify patterns, and generate outputs. While higher parameter counts do not automatically translate to superior capability, they remain a primary benchmark for total compute scale and memory capacity.

Crossing the Trillion-Parameter Threshold

Led by former Google DeepMind scientist Wu Yonghui, ByteDance’s 2,000-person Seed AI team is executing a massive infrastructure push that far outpaces domestic competitors:

  • Moonshot AI: Its recently launched Kimi K3 model stands at 2.8 trillion parameters.
  • Meituan & DeepSeek: Previous domestic leaders, including Meituan’s LongCat-2.0 and DeepSeek’s V4-Pro, hovered around 1.6 trillion parameters.
  • ByteDance: The new 10-trillion-parameter architecture is more than three times the size of its closest domestic rival.

On the global stage, Western frontier labs rarely publish exact parameter counts. However, industry estimates place Anthropic’s flagship Mythos 5 model at roughly 8 trillion parameters and Fable 5 at approximately 5 trillion. If those projections hold, ByteDance’s system will match or exceed the scale of top American systems.

Independent Development over Distillation

Unlike competitors that leverage distillation, training smaller models on the outputs of existing US systems, ByteDance founder Zhang Yiming has instructed teams to build foundation models independently. The effort also bypasses open-weights releases, keeping the architecture closed to maximize proprietary frontier capabilities.

The model is currently in early pre-training, a compute-heavy phase expected to take three to six months. Following pre-training, the system will move to fine-tuning, safety testing, and post-training alignment before any commercial integration into ByteDance’s ecosystem, which includes its Doubao chatbot and Volcano Engine cloud platform.

Topics

The Morning Brief

Start every day knowing what matters in tech

Six stories chosen by our editors, in your inbox at 06:00 UTC. Free.

Free, one email each morning. No tracking pixels, and unsubscribing takes one click.

See more Blockrora in Google

Make us a preferred source and our stories show up more often in Top Stories.

Add as preferred source

Founder and editor

Founder and Editor at Blockrora. A 20-year veteran of fintech, Web3 ecosystems, and software development, Warren writes on AI infrastructure, decentralised networks, and market reality checks. Trail walker and coffee purist.

More from Warren →
All Latest News →

Keep listening, free

Audio comes with the Morning Brief, our free daily email of the stories that matter. Sign up once and every story is yours to listen to on this device.

Free, one email each morning. No tracking pixels, and unsubscribing takes one click.

Already get the Brief? Enter your email to unlock this device.

Search Blockrora The Morning BriefToday’s stories, every morning ListenEvery story, read aloud