Sapient trains 1B-parameter foundation model from scratch for about $1,500

Researchers at Sapient trained a 1B-parameter language model from scratch for about $1,500, VentureBeat reported. The model, HRM-Text, replaces standard Transformers with a Hierarchical Recurrent Model that decouples computation into slow-evolving strategic and fast-evolving execution layers, and trains exclusively on instruction-response pairs instead of next-token prediction over raw web text. Sapient said the model achieved performance competitive with much larger open models on key industry benchmarks, at a fraction of the cost and tokens of normal LLMs. "Enterprises today face three compounding problems: training is expensive, infrastructure is heavy, and experimentation cycles are too slow," Sapient Intelligence CEO Guan Wang said.