ByteDance is reportedly in the middle of training an artificial intelligence model carrying as many as 10 trillion parameters, a figure that would place the system in the same weight class as Anthropic’s flagship Mythos technology, according to a Financial Times report published Friday that cited people familiar with the effort.
Should the model ship at that size, it would exceed by more than threefold the largest system yet produced in China: Kimi K3, the model released by startup Moonshot AI, which carries 2.8 trillion parameters. Kimi K3’s debut reset the domestic benchmark. Prior to its arrival, the top of the Chinese market belonged to Meituan’s LongCat-2.0 and DeepSeek’s V4-Pro, each built with 1.6 trillion total parameters. Several additional Chinese developers have since cleared the one-trillion-parameter mark, making the threshold less of a milestone than it was a year ago.
Parameters, for context, are the numerical values a model derives from its training data — the settings it tunes in order to spot patterns, produce responses and complete assigned tasks. The industry leans on parameter counts as convenient shorthand for how large a system is, but the number describes scale rather than skill. A bigger model is not automatically a better one.
Measuring Chinese models against their American counterparts is an inexact exercise for a simpler reason: Anthropic and OpenAI decline to publish parameter counts for their frontier systems, leaving Fable, Mythos and GPT-5.5 as effectively unmeasured quantities in public. What exists instead are estimates. The Financial Times reported that industry analysts put Anthropic’s most advanced system, Mythos 5, in the neighborhood of 8 trillion parameters, with Fable 5 around 5 trillion. Against those figures, a 10-trillion-parameter ByteDance model would sit near — and by raw count, above — the Mythos tier.
The system is currently in pre-training, the report said. That phase generally runs three to six months, after which a model still requires fine-tuning before any public release. A launch date, in other words, is not imminent.
The reported project lands amid a broader acceleration among Chinese technology companies, which have compressed their release schedules to stay in step with the global race. The pressure runs in two directions at once: developers on both sides of the Pacific are chasing more capable systems while trying to keep inference costs from climbing to the point where the models become impractical to operate at scale.
