Home / Technology / AI Costs Slashed: New Model Trains for $1,500
AI Costs Slashed: New Model Trains for $1,500
11 Jun
Summary
- Sapient's HRM-Text trains models from scratch affordably.
- Model achieves competitive performance with larger LLMs.
- Training cost is approximately $1,500 on 16 GPUs.
Sapient has introduced HRM-Text, an AI architecture designed to overcome the prohibitive costs and resource demands of training foundation LLMs from scratch. Traditional methods relying on internet-scale data and immense computational power are being challenged by Sapient's hierarchical recurrent model (HRM).
HRM-Text decouples computation into strategic and execution layers, training exclusively on instruction-response pairs instead of brute-force next-token prediction. This method proved effective in training a 1-billion parameter model for approximately $1,500 on 16 GPUs, achieving performance on par with larger models.
This breakthrough makes foundational LLM pretraining accessible beyond highly resourced institutions. Enterprises can now affordably develop specialized reasoning models, potentially pairing them with external knowledge stores for tailored applications. The efficiency demonstrated suggests a future where AI development shifts from an infrastructure challenge to a strategic business decision.
