ByteDance has started pre-training a model that three unnamed sources describe as reaching roughly 10 trillion parameters, according to a KuCoin News flash that cites a ChainCatcher report attributed to the Financial Times. If the number holds, the model would land near the scale industry analysts already assign to Anthropic’s largest system. The claim travels through two layers of secondhand sourcing before reaching readers, which is worth weighing before treating it as confirmed.
Parameter count stopped being the metric labs compete on for bragging rights some time ago. The industry moved toward benchmark scores, inference cost, and how reliably a system completes multi-step tasks. A 10-trillion figure, unverified at this stage, functions as much as a positioning statement aimed at competitors and investors as it does an engineering disclosure, especially before the model has produced a single public result.
For scale, the report puts the new system at about three times the size of Kimi K3, the largest released model from Moonshot AI, the Beijing-based lab behind the Kimi line. Anthropic does not publish parameter counts for its own systems, but analyst estimates cited in the same report place Mythos 5 near 8 trillion parameters and Fable 5 near 5 trillion. None of those three figures, ByteDance’s, Mythos 5’s, or Fable 5’s, come from the companies that built the models. They are outside approximations, passed along through a single wire item rather than disclosed directly.
The model remains in pre-training, per the report, a stage that commonly runs three to six months before a lab moves to fine-tuning and eventually a release decision. That timeline pushes any real evaluation of the claim months into the future. There is no benchmark to check, no red-teaming result to read, and no confirmed final parameter count, because none of that exists yet at this stage of development. ByteDance has not commented publicly on the training run, and the report does not indicate when, or whether, the company plans to.
The broader context matters more than the single number. Chinese labs, Moonshot AI and DeepSeek among them, have shipped large open and closed models at a steady pace over the past two years, and each release narrows a gap that used to favor U.S. frontier labs by default. A 10-trillion-parameter training run, even unconfirmed, fits that pattern: it signals where ByteDance wants to be measured against Mythos 5 and Fable 5, regardless of what the model can actually do once training finishes.
Treat the figure as a claim in circulation rather than a specification until ByteDance or a primary outlet confirms it independently. Operators tracking Chinese open-weight competition should watch for confirmation that the model has moved into fine-tuning, which the reported timeline would place between the fourth quarter of 2026 and the first quarter of 2027, rather than reacting to the parameter count on its own.
Reported by KuCoin News on August 7, 2026, citing a ChainCatcher report attributed to the Financial Times.
