AI
ByteDance Pushes AI Limits With 10 Trillion-Parameter Model to Challenge US Leaders
The Chinese tech firm is training a model significantly larger than anything it has released before, signaling intensifying competition between Beijing's AI labs and top American rivals.
Ars Technica
ByteDance is developing what could become one of the world's largest AI models as part of a broader effort by Chinese technology firms to match—and potentially exceed—the capabilities of leading US research labs.
The company is currently in the early stages of pre-training a model estimated at up to 10 trillion parameters, sources familiar with the project revealed. This would represent a dramatic leap: the figure dwarfs Moonshot's Kimi K3, previously the most capable model to emerge from China, which measures around 3 trillion parameters.
Where ByteDance Stands#
ByteDance's model remains in its initial training phase, which typically spans three to six months before the system undergoes fine-tuning and potential commercial release. The team has not yet locked in a final parameter count, with that determination expected later in development. Even so, the scale being explored reflects the company's determination to push past existing thresholds.
The biggest US models remain larger in disclosed estimates—Anthropic's Mythos 5 is thought to contain roughly 8 trillion parameters, while its Fable 5 variant comes in around 5 trillion—though Anthropic typically keeps precise specifications confidential.
The Broader Race Heats Up#
Recent releases from competitors like Moonshot and Alibaba have demonstrated marked improvements in benchmark performance, with some metrics placing them nearly on par with Anthropic's Fable 5. Meanwhile, Anthropic's most advanced Mythos 5 remains restricted to approved users following a temporary suspension announced in June over security considerations.
Multiple Chinese research outfits are reportedly developing systems of comparable scale to Fable 5, but ByteDance's push toward the absolute frontier makes it the most aggressive player in the current competition.
What Size Actually Means#
While parameter count indicates a model's fundamental storage and processing capacity, raw numbers don't tell the whole story. Training methodology, data quality, and architectural innovations all meaningfully influence actual performance and capability. ByteDance's ambitions suggest the company believes scale, paired with its engineering resources, will help it compete at the highest tier of AI development.
