ByteDance Tests AI Model That Could Match Anthropic’s Top Systems

ByteDance’s Ambitious Blueprint for a 10-Trillion-Parameter Model

ByteDance, the parent company of TikTok, is in the early pre-development phase of a new artificial intelligence model that could scale to 10 trillion parameters, according to three people familiar with the project cited by the Financial Times. If realized, the system would be three times larger than China’s current largest model, Moonshot’s Kimi K3, and comparable in raw size to the most advanced closed systems from Anthropic, whose Mythos 5 model is estimated by industry analysts to contain around eight trillion parameters.

The project is still in its infancy; sources say the pre-development work will take three to six months, after which fine-tuning and a final decision on the exact parameter count will follow. The company, led by founder Zhang Yiming, has been quietly pouring resources into AI infrastructure. Over the past three years, ByteDance has invested heavily in data centers, semiconductors and global talent for its Seed AI research unit, which now numbers about 2,000 professionals worldwide.

The move reflects the growing ambition of Chinese AI labs to close the gap with US leaders. ByteDance already operates Doubao, China’s largest chatbot with 324 million monthly active users, and its SeeDance video-generation model is considered among the most advanced globally. Unlike rivals such as Meta and DeepSeek, which have open-sourced key models, ByteDance has kept its systems proprietary—a stance that could have implications as US authorities prepare to tighten controls on closed AI models while leaving open-source models largely untouched.

What a ByteDance AI Breakthrough Means for the Global Tech Order

A Parameter Arms Race—With Plenty of Caveats

A 10-trillion-parameter model would mark a step-change in China’s AI scale, dwarfing the domestic record holder and leaping directly toward the frontier defined by Anthropic and arguably OpenAI’s rumored systems. Parameter count alone, however, does not guarantee superior performance; real-world effectiveness depends equally on training data quality, architecture efficiency and fine-tuning—areas where US firms still hold considerable leads. Still, the sheer ambition signals that ByteDance has both the capital and the institutional will to compete at the very top of the model-size ladder.

ByteDance’s Closed-Door Strategy in a Fragmented Landscape

ByteDance’s decision to keep its AI models proprietary sets it apart from most Chinese peers. Moonshot and Alibaba have publicly demonstrated competitive performance on benchmarks and have embraced open-source releases in a bid to build developer ecosystems. A closed approach gives ByteDance more control and could protect proprietary advantages, but it also raises the stakes. If the model succeeds, it could become a powerful wedge for ByteDance’s existing consumer platforms. On the flip side, a closed model may face stricter US export controls than open systems—an emerging regulatory dividing line that already favors open-source creators like Meta and DeepSeek.

The Geopolitical Dimension: Chips, Controls and Catch-Up

ByteDance’s AI push is unfolding against a tightening US export regime on advanced semiconductors and AI safety rules. The Biden administration’s framework, reported in US media, is expected to impose new restrictions on proprietary models from OpenAI, Google and Anthropic while allowing open models to be downloaded and modified freely. If that approach crystallizes, ByteDance’s closed system could fall squarely into the restricted category, complicating its ability to serve global customers and potentially limiting access to cutting-edge hardware from Nvidia. Still, the investment spree in data centers and recruitment suggests ByteDance is racing to build momentum before any broader clampdown.

What This Means for Anthropic and the US AI Lead

Anthropic’s perceived lead in safe, high-capability models has been a cornerstone of its commercial appeal and fundraising narrative. A credible Chinese rival with comparable raw scale—backed by one of the world’s richest app ecosystems—would challenge the assumption that the most advanced AI will remain a US-centric asset. Even if ByteDance’s model launches a year or more from now, the psychological effect on the market, and on enterprise buyers weighing AI vendors, could be significant. For now, however, the project remains speculative, with no guarantee it will reach full training or match the safety and reliability standards that Anthropic has made its hallmark.

Playing the Next Move: Competitive and Strategic Takeaways

• For US AI developers: Monitor ByteDance’s model development as an indicator of China’s rapid progress in closed-model capabilities. Even a moderately performing system at this scale could erode the perceived uniqueness of US enterprise offerings, especially in price-sensitive markets. ByteDance’s existing user base via Doubao and TikTok provides a ready distribution channel.
• For chipmakers and investors: ByteDance’s ongoing data-center buildout and recruiting surge are tangible signals of sustained demand for AI hardware, even as export controls evolve. A 10-trillion-parameter training run would require enormous compute, potentially benefiting Nvidia and its Asian partners despite restrictions.
• For policymakers: The regulatory dichotomy between open and closed models is becoming a key strategic variable. If US rules impose heavier burdens on proprietary systems, homegrown closed models from China may find it harder to compete internationally—but could also drive more rapid innovation in open-source AI inside China, as seen with DeepSeek.
• For enterprises evaluating AI suppliers: Keep an eye on ByteDance’s eventual model announcements, but do not equate parameter count with enterprise readiness. Performance in real-world tasks, safety, and integration support will matter far more. The anticipated 2027 launch window means near-term procurement decisions are unlikely to be disrupted.