ByteDance is training an AI model with as many as 10 trillion parameters, according to the Financial Times. That size would put it close to Anthropic's most advanced system. Anyone tracking the global AI race should watch this, because it signals how fast Chinese labs are closing the gap with US leaders.
What actually happened
The model is three times larger than Moonshot's Kimi K3, the biggest Chinese model released so far, according to three people with knowledge of the matter, per Ars Technica. Pre-training typically takes three to six months, and the final size will be set later. Anthropic does not disclose parameter counts, but industry estimates put Mythos 5 at about 8 trillion parameters and Fable 5 at about 5 trillion. Mythos 5 has been limited to approved organizations since a June security ban. ByteDance's consumer app Doubao has 324 million monthly active users, the most in China. Its Seed research team, led by former Google DeepMind scientist Wu Yonghui, has about 2,000 members. Founder Zhang Yiming told the Seed team, in an internal meeting two weeks ago, to target "world-leading model capabilities" without worrying about short-term setbacks.
How we got here
Chinese labs have gained ground quickly this year. Models from Moonshot and Alibaba have posted benchmark scores that trail only Anthropic's Fable 5 in some areas. ByteDance has kept a lower public profile than rivals, keeping most of its models closed while pouring resources into its Volcano Engine cloud unit and custom AI chip plans. Unlike some competitors, ByteDance's Seed team has avoided distilling knowledge from other labs' models for more than a year, a slower but more independent path. That approach has drawn internal debate over whether it has cost ByteDance time in the race to match Western frontier labs.
Why this matters for you
For builders, a larger ByteDance model could mean fiercer competition for developers choosing between Chinese and US AI platforms. For users, more capable Chinese models could lower the cost of accessing advanced AI, especially if released openly. For AI infrastructure providers and chipmakers, ByteDance's push adds another buyer racing for compute and custom silicon. Access matters too, Mythos 5 remains restricted to approved organizations, so much of the frontier stays gated. If ByteDance's model performs well, expect faster iteration cycles across the industry and more pressure on Anthropic and its peers to match training scale rather than efficiency alone.
The bigger question
If ByteDance's model matches Anthropic's scale, does raw parameter count still determine which AI systems lead the world, or has data quality and training method become the bigger factor? The report notes both matter, but companies keep racing for size. As more labs approach 10 trillion parameters, the industry may need a clearer answer about what actually separates a leading model from a merely large one.
What to watch
ByteDance's pre-training phase typically runs three to six months, meaning the model's fine-tuned form and eventual release, if it happens, could land in late 2026 or early 2027. Watch for confirmation of final parameter count, benchmark comparisons against Anthropic's Mythos 5 and Fable 5, and whether ByteDance opens or keeps the model closed. Bonuz will track how this affects the hardware and chips powering next-generation AI devices, including wearables.
