The number that matters is 10 trillion: the parameter count of the model ByteDance is now pre-training, per three people familiar with the project speaking to the Financial Times. If the figure holds through training, it would be the largest model ever attempted in China, three times Moonshot's Kimi K3, released last week at 2.8 trillion and currently the biggest Chinese model released. It would also land near the industry's unconfirmed estimate for Anthropic's Mythos 5, around 8 trillion parameters, a figure Anthropic has never published and does not confirm. 'Reported by the FT' is the load-bearing phrase in this paragraph, and it stays attached to everything below.
The details the insiders gave: the model sits in pre-training, a phase that typically takes three to six months, inside ByteDance's Seed team, a 2,000-person unit that founder Zhang Yiming has internally told to aim at long-term world-leading model capability. One person added that ByteDance has gone more than a year without distillation, the practice of training a model on another lab's outputs, which reads as a direct answer to the distillation accusations the Chinese labs traded with the US labs this spring. The FT's report also notes that xAI is training Grok variants at six and ten trillion parameters on its Colossus 2 cluster, per Elon Musk, so the size race is running again on at least two clusters on two continents.
The honest reading requires two caveats, one technical and one editorial. Technical: parameter counts set capacity, not performance; data quality, training method and post-training decide what the model can do, and a well-trained smaller model beats a sloppy giant every week of the year. Editorial: ByteDance has announced nothing, and everything known is three anonymous insiders plus the FT's track record, which is good but not an announcement. What makes it more than rumor is the direction of travel: Kimi K3 at 2.8 trillion shipped open last week, and if ByteDance follows its usual pattern with Seed models, the closed-lab lead on raw size keeps narrowing.
For builders the watch items are practical. Whether the model ships at all, pre-training runs get canceled quietly; whether it ships open, which is where ByteDance's size would actually reach you; and whether it lands near Mythos 5 in capability, in which case the open-weight ceiling moves again. The one thing not to do with this story is what everyone will do with it: read '10 trillion' as a capability claim. It is a budget claim.