On July 23, ByteDance's Seed Edge team officially launched the Seed STEM Scientist program, planning to recruit 100 STEM scholars globally, providing compute resources, model experience, and AI research team support, driving AI-accelerated scientific discovery around real research problems. The first phase of the program runs about 6 months, applications close September 30, and the work location is in Beijing Haidian. This is the first time the ByteDance Doubao large-model matrix has embedded base-model supply into the academic chain in the form of a research residency + open compute, intersecting head-on with the AI4Science path of China's top AI labs. Program design: from giving models to giving desks Unlike simply open-sourcing models or opening APIs, the Seed STEM program requires selected scholars to enter ByteDance's Beijing offices as scientist advisors or PhD interns, collaborating with the Seed team for 6 months. The selection threshold focuses on three hard criteria: a PhD in a STEM field or equivalent research level, daily deep use of AI, and a preference for those who can program fluently. In other words, ByteDance isn't looking for occasional API users, but for peers willing to embed AI into their own research pipeline and treat the model as a collaborator. Beyond ample compute, the package also includes an industry-competitive salary — meaning ByteDance is using real money to buy a high-quality feedback loop. Why Seed Edge, not Doubao? Seed Edge is ByteDance's long-term internal research program, positioned to explore the unknown boundaries of intelligence, clearly differentiated from the commercial main line of Doubao/Jimeng. Putting the STEM program under Edge rather than the Doubao brand means ByteDance sees this round of research collaboration as frontier exploration, not pre-launch user research. This positioning also means the Seed Edge team has more freedom in compute scheduling — 6-month on-site research can't be satisfied with ordinary user-level quotas, which is a signal to both budget approval and cluster management. Gap and difference with Anthropic, Google Compared with Anthropic packaging its AI scientist workbench as a product (Claude Science beta is already online), or Google bringing Nobel laureate John Jumper on board to do Protein LLM, ByteDance's path is closer to DeepMind's early model: put the model team and the research team in the same room physically, iterate as you go. This model's strengths are fast feedback and engineering-ready output; its weakness is hard-to-scale and long PhD-training cycles. 100 scholars, 6-month cycle — ByteDance is essentially betting on a small-and-deep research organization paradigm. A judgment on the AI4S track The real signal of the Seed STEM program isn't how many PhDs it hired, but that it's willing to open up compute and model scheduling power to academic teams. In China, the research institutions that have consistently built up base-model supply are few — DeepMind has AlphaFold, Anthropic has Constitutional AI, and the ByteDance Doubao matrix has Seed Edge. When a Chinese vendor is willing to bear the cost of a research residency, it means the differentiated competition of base models has quietly shifted from leaderboard scores to who can get the model into scientists' workflows faster.