rotative
the decoder · August 7, 2026

Bytedance trains 10T params to beat Kimi K3

bytedance is training a 10 trillion parameter model, dwarfing moonshot's kimi k3, signaling a fierce new front in the us-china ai arms race.

read in 简体中文 · 繁體中文 · 日本語 · 한국어

bytedance is pretraining a massive ai model with up to 10 trillion parameters, according to financial times sources, which would triple the capacity of china's current largest model, moonshot's kimi k3. the project utilizes a 2,000-person team called seed, with founder zhang yiming explicitly setting a goal for world-leading model capabilities over the long term. insiders claim the company has avoided distillation for over a year to ensure higher training quality and originality. this initiative places bytedance in direct competition with top us labs like anthropic and xai, which are reportedly developing models of similar scale on massive gpu clusters.

read at source ↗

← rotative