ByteDance's large-model department Seed has completed a new round of organizational restructuring, setting up four new first-level departments under its foundation model team, and has started pre-research on a next-generation model with more than 5 trillion parameters, reports said on November 16.
On the capital side, ByteDance completed a 29.6 billion US dollar syndicated loan in early September, the second-largest syndicated transaction in Asia this year, with funds mainly directed at data center expansion, AI chip procurement and infrastructure.
On the pressure side, Seed 2.0 ranked ninth globally on the LMSYS Chatbot Arena in February as the only Chinese model in the top ten at the time, but Seed 2.1 released in June has since been overtaken on core benchmarks by DeepSeek V4, Kimi K3 and GLM-5.2.
Zhang Yiming said in an internal meeting that data is the core bottleneck, while MIIT-affiliated expert Pan Helin argued ByteDance's real weakness is compute, noting that data must be converted into capability with sufficient computing power. The divergence highlights the dilemma facing domestic foundation models as leaders return to the most expensive path of scaling base models.