sapientinc ORG

Hierarchical Reasoning Model Official Release

A tiny 27-million-parameter AI model that uses two cooperating thinking modules, slow abstract planning and fast detailed computation, to solve complex reasoning tasks like Sudoku and mazes better than much larger models.

HRM-Text is a 1B text generation model based on the HRM architecture, strengthened by task completion and latent space reasoning.

Pretraining stack for a 1B-parameter hierarchical recurrent language model that can be trained from scratch on H100 GPUs for about $1000 of rental compute.

Data pipeline for HRM-Text pretraining

Four-stage data preparation pipeline that cleans Hugging Face datasets into instruction-response pairs, tokenizes them in Rust, and samples balanced epochs for HRM-Text pretraining.