ByteDance Is Pre-Training a 10-Trillion-Parameter AI Model — Here’s What We Know

Reuters could not independently confirm the Financial Times report .

How It Compares to Anthropic's Mythos 5 and Other Major Models

ByteDance's 10-trillion ceiling would place it near or slightly above Anthropic's Mythos 5, which industry estimates peg at approximately 8 trillion parameters . Here is how the known landscape looks:

Raw parameter count, however, is not a direct proxy for capability — it is a rough architectural measure.

The Critical Caveat: Total Parameters vs. Active Parameters

The reported 10 trillion "total parameters" refers to the full parameter count of the model's architecture. Many state-of-the-art models use Mixture-of-Experts (MoE) architectures, where only a fraction of parameters are activated per inference. For example, a 10-trillion total parameter model might have only 1–2 trillion active parameters at any given time. The FT report and subsequent coverage have not clarified whether the 10 trillion figure represents total parameters, active parameters, or a mix. This distinction is critical for fair comparison with other models .

Zhang Yiming's Anti-Distillation Directive

ByteDance founder Zhang Yiming has told the Seed team that the company will not use AI distillation on rival model outputs to build this model, even if it means falling behind domestic competitors in the short term . Reports indicate he reiterated this position at a management meeting around August 6, 2026, framing it as a long-term strategy of "delayed gratification" and independent capability-building . The decision is partly influenced by ByteDance's complex regulatory relationship with the US government (over TikTok) and the desire to avoid legal and export-control exposure .

Distillation is widely used among Chinese AI labs to rapidly mimic US frontier models — Zhang's directive is a notable strategic departure .

ByteDance's Seed Team: Size and Leadership

ByteDance established the Seed division in early 2023 in direct response to ChatGPT's emergence . Key details about the team responsible for this 10-trillion-parameter model:

• Size: The Seed team employs roughly 2,000 people across China, including researchers, infrastructure engineers, and data labeling staff .
• Research head: Wu Yonghui — a former Google DeepMind VP of Research and Google Fellow, who joined ByteDance in February 2025 .
• Foundation model project lead: Xiang Liang (项亮), head of Seed Foundation, collaborating with pre-training data lead Shen Ke (沈科) .
• Reporting structure: Wu Yonghui reports directly to ByteDance CEO Liang Rubo .
• Reorganization: In 2025, ByteDance folded its AI Lab and robotics team into Seed, consolidating all foundation model R&D under one roof .

US-China AI Race Dynamics

ByteDance's 10-trillion-parameter project is widely seen as a direct bid to close the gap with US frontier labs — particularly Anthropic and OpenAI . The broader context includes:

• The US government has imposed export controls on advanced AI chips to China, forcing ByteDance to rely on domestic and stockpiled hardware .
• ByteDance's massive infrastructure buildout for this model is partly a "paramilitary-style scale-up" driven by Zhang Yiming's personal involvement in reviewing computing resource allocation, according to TechTimes .
• The project comes as China's AI sector has been energized by DeepSeek R1 (January 2025), which demonstrated that Chinese labs can achieve frontier-competitive results despite chip restrictions .
• Anthropic's Mythos 5 has been caught up in US national security debates — its full cybersecurity and biodefense capabilities were briefly suspended by the US government in June 2026 before being partially restored for "trusted" organizations . This highlights how frontier models themselves are becoming strategic assets in US-China competition.

ByteDance's Closed-Source Approach and Product Lineup

ByteDance has not released details on whether this 10-trillion model will be made public. However, the company's general posture has been closed-source and product-integrated:

• ByteDance does not open-source its flagship large models, unlike Meta (Llama) or some Chinese rivals like DeepSeek.
• The company's consumer-facing AI brand is Doubao (豆包), which integrates ByteDance's Seed-series foundation models into chat, assistant, and content-generation products .
• ByteDance's 2026 AI priorities, outlined by Wu Yonghui, include: world models (targeting parity with Google's Genie 3 by end of 2026), a Seedance coding agent, and continued Doubao commercialization .
• The 10-trillion model, if completed, would likely form the backbone of Doubao and other internal products rather than being released as a separate open model.

In short: ByteDance is making an enormous, long-term bet on independent frontier-scale AI, shaped by Zhang Yiming's personal directive to avoid shortcut distillation, led by a 2,000-person Seed team under Wu Yonghui, and fueled by the intensifying US-China technology rivalry. The project is in early pre-training, with significant uncertainty about final parameter count, timeline, and capability outcomes.