Researchers say they trained a foundation model from scratch for about $1,500

Sapient’s HRM-Text replaces Transformers with a hierarchical recurrent model trained only on instruction-response pairs, cutting pretraining cost and data needs. A 1B-parameter model trained on 40B tokens reportedly matched or beat larger open models on key reasoning benchmarks, costing about $1,500 and 1.9 days on 16 GPUs.