To date, only two entities worldwide have publicly applied Recursive Self-Improvement (RSI) to frontier pretraining from scratch: Recursive, a U.S. star startup valued at $4 billion, and the team behind Shanghai's Classroom 801. This team has jointly released the first complete public technical report on pretraining-level RSI, the 801-30B base model, and high-quality lightweight SFT data—training a model that ranks among the top tier of open-source base models of the same size in just three months. The team decomposed the traditional grand RSI loop—which is difficult to close directly—into three types of small loops that can be closed quickly, and leveraged shared production memory to accumulate experience, solving multiple cross-team challenging problems and significantly improving training efficiency and model performance. Evaluations show that the model performs outstandingly in multi-dimensional comparisons, capable of competing head-on with fully post-trained models from multiple vendors using only lightweight SFT. Implementing RSI at the pretraining level enables AI to participate in the core process of manufacturing AI, driving a transformation in how base models are produced, achieving compounding growth of pretraining experience, and is particularly critical for compute-constrained entities seeking to improve intelligent output per unit of compute.