Researchers Develop Low-Cost Foundation Model Training Method at $1,500

Sapient Intelligence researchers have created HRM-Text, a 1-billion-parameter foundation model trained for roughly $1,500—a dramatic reduction from the millions typically spent on foundation model development. The key innovation is replacing standard Transformer architectures with a Hierarchical Recurrent Model (HRM) that decouples computation into strategic and execution layers, allowing the model to train exclusively on instruction-response pairs rather than raw internet text. This approach addresses what the researchers identify as computational inefficiency in current large language models, which spend significant resources on tasks like reconstructing prompts that are already known at inference time. According to Sapient CEO Guan Wang, the method addresses enterprise pain points: expensive training, heavy infrastructure requirements, and slow experimentation cycles. The model achieved competitive performance on key industry benchmarks while using a fraction of the tokens required by standard LLMs, potentially enabling organizations to develop proprietary reasoning models without relying on external frontier models or sending sensitive data to third parties.