ByteDance is pre-training a 10-trillion-parameter model, per FT

ByteDance is pre-training a model of up to 10 trillion parameters, the Financial Times reports — roughly three times Moonshot's Kimi K3 at 2.8 trillion, the largest model released in China, and above the ~8 trillion FT attributes to Anthropic's Mythos 5. The report is unconfirmed: Reuters could not verify it, ByteDance declined to comment, and final size, fine-tuning, and release are undecided. If the number holds after the three-to-six-month pre-training run, the largest frontier training job would sit at a Chinese consumer-app company rather than a US lab.

Source: mobileworldlive.com ↗