ByteDance Is Training a 10-Trillion-Parameter AI Model

ByteDance has quietly kicked off training on a monster: an AI model with roughly 10 trillion parameters. It’s a bold move that throws the company into the center of a global arms race over model scale and capability.

Pretraining is underway, according to people familiar with the matter. That stage—where the model ingests massive amounts of raw data to learn patterns—typically runs three to six months before fine-tuning and public release. Parameters are the internal settings a model uses to encode knowledge; they don’t tell the whole story, but they are a widely used shorthand for a model’s size and potential.

Why does scale matter? Bigger models can capture more nuance and generalize better across tasks, although returns aren’t linear and efficiency still counts. ByteDance’s reported 10-trillion target would make this model more than three times larger than Kimi K3, a 2.8-trillion-parameter system from a Chinese startup, and far larger than earlier domestic entrants such as LongCat-2.0 and V4-Pro, each in the 1.6-trillion range.

Direct comparisons with leading US-made systems remain tricky. Firms like Anthropic and OpenAI seldom publish exact parameter counts for their flagship models. Independent estimates place Anthropic’s advanced offering, Mithos 5, near eight trillion parameters and Fable 5 close to five trillion. If those estimates hold, ByteDance’s new effort would put it in territory comparable to Anthropic’s most sophisticated systems.

There’s another layer to this story: logistics and cost. Training at this scale is expensive and energy-hungry. Chinese tech companies are racing to compress development cycles and squeeze more efficiency out of their stacks, aiming to field powerful systems without runaway operational costs. It’s a pragmatic answer to an unforgiving market.

Whether sheer parameter count will decide the winners is still an open question, but ByteDance’s push underscores how the competitive frontier has shifted—from clever research papers to industrial-scale engineering and deployment.

Expect more announcements and incremental improvements rather than a single, defining breakthrough. The real battle will be over how these systems are tuned, scaled, and integrated into products that people actually use.