ByteDance Trains 10 Trillion Parameter AI to Rival Anthropic
ByteDance is training an AI model that could reach 10 trillion parameters, putting it near Anthropic's biggest systems. That scale shows how quickly Chinese AI labs are closing the gap with US frontrunners, a shift anyone watching the AI race should note.
What actually happened
Three people close to the project told Ars Technica the model is triple the size of Moonshot's Kimi K3, currently China's largest released model. Pre-training normally runs three to six months, so ByteDance has not fixed a final parameter count yet. Anthropic keeps its own figures private, but industry estimates put Mythos 5 near 8 trillion parameters and Fable 5 near 5 trillion, according to ByteDance Trains Massive 10 Trillion Parameter AI Model. Mythos 5 has stayed limited to vetted organizations since a security-related ban in June. ByteDance's Doubao app already leads China with 324 million monthly users. Its Seed research unit, led by former Google DeepMind scientist Wu Yonghui, has around 2,000 staff. Founder Zhang Yiming reportedly told the team two weeks ago to chase world-leading model capabilities and not worry about short-term setbacks.
How we got here
Chinese AI labs made fast progress through 2025. Moonshot and Alibaba systems have posted benchmark scores trailing only Fable 5 in some tests. ByteDance stayed quieter than rivals, keeping most releases closed while funding its Volcano Engine cloud arm and in-house chip work. Its Seed team has skipped distilling knowledge from other labs' models for over a year, a slower but more independent route. That choice has stirred internal debate over whether it cost ByteDance time.
Why this matters for you
For app builders, a larger ByteDance model raises the stakes in choosing between Chinese and US AI platforms. For users, stronger Chinese models could lower the cost of advanced AI, especially with open releases. Chipmakers and cloud providers gain another major buyer chasing compute and custom silicon. Access still matters, since Mythos 5 remains gated to approved partners. A strong ByteDance result could speed up release cycles industry-wide and push rivals toward scaling training further.
The bigger question
If ByteDance's model matches Anthropic's scale, does parameter count still decide which AI system leads, or do data quality and training method matter more? Labs keep racing for size even as that question stays open. As more models near 10 trillion parameters, the industry may need a clearer way to define what makes a model leading rather than merely large.
What to watch
Pre-training typically takes three to six months, so a finished version could land in late 2026 or early 2027, if ByteDance releases it at all. Watch for the confirmed parameter count, benchmark comparisons against Mythos 5 and Fable 5, and whether the model ships open or closed. bonuz.market will track how this affects chips and devices behind next-generation wearables.
