ByteDance is training an AI model that could approach Anthropic’s cutting‑edge Mythos system, as Chinese firms narrow the gap with top US labs. The Chinese tech giant is training a model with up to 10 tn parameters — three times larger than Moonshot’s Kimi K3, the largest Chinese model to date, according to three people.

The ByteDance model is pre‑trained for three to six months before fine‑tuning and release. While Anthropic keeps model size secret, estimates put Mythos 5 at about 8 tn parameters and Fable 5 at 5 tn. Parameter count sets memory limits, but capability also depends on data quality and training methods.

ByteDance’s push shows Chinese labs aim to outperform US peers in the most advanced AI level. Recent Moonshot and Alibaba models rank near Anthropic’s Fable 5 on benchmarks. Mythos 5, only available after a June security ban, is being targeted by multiple Chinese labs, with ByteDance leading the effort.

ByteDance, parent of TikTok, keeps a low profile, with its See Dance video model and Doubao, a popular Chinese app with 324 mn monthly users. Over three years it has heavily invested in AI, built data‑centre capacity, expanded Volcano Engine, and pursued custom chips, while its Seed team of about 2,000 works independently, a strategy championed by founder Zhang Yiming.