TikTok's parent company is reportedly building one of the largest AI models ever attempted.
According to Ars Technica, ByteDance is training a model with 10 trillion parameters in a bid to rival Anthropic. Mezha, in coverage aggregated by Google News, reports that the figure would put ByteDance ahead of Anthropic's Mythos model.
To put the number in context, Livemint notes that if the reported figure is accurate, the ByteDance model would be more than three times larger than Kimi K3 from Chinese startup Moonshot AI, which is built with roughly 2.8 trillion parameters.
Parameters are the internal values a model adjusts as it learns — loosely, the dials that encode what it knows. More parameters generally means more capacity to absorb patterns from training data, and it also means far more expensive computing hardware and electricity to train and run the thing. Parameter count is not a direct measure of how good a model is: a smaller, better-trained model can beat a larger, sloppier one. But at the frontier, scale has historically tracked with capability, which is why the number draws attention.
It is worth being precise about what has and hasn't been established here. The reporting describes a model in pre-training, not a released product, and Livemint frames the parameter count conditionally — "if the reported figure is accurate." No benchmark results, release date, or pricing have been reported in these sources, and there is no public comparison yet against specific Anthropic or OpenAI systems.
Why it matters: if ByteDance is genuinely pre-training at 10 trillion parameters, it signals that the most expensive tier of AI development is no longer the preserve of a handful of US labs — and that the company best known for a short-video app is now spending like a frontier AI lab.