sapientinc/HRM

Health v2 · maintenance only

Flags: no_releases

Adoption not part of the score

12619 stars · 1825 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

Official PyTorch implementation of the Hierarchical Reasoning Model (HRM), a 27M-parameter recurrent architecture with high-level and low-level modules that performs sequential reasoning in a single forward pass without chain-of-thought data. It achieves strong results on Sudoku, maze pathfinding, and the ARC benchmark with only ~1000 training samples and no pre-training.

Use cases

• reproduce HRM results on Sudoku and maze reasoning tasks
• train a small reasoning model on the ARC benchmark
• experiment with brain-inspired hierarchical recurrent architectures
• research latent-space reasoning without chain-of-thought supervision
• benchmark small models against large LLMs on reasoning tasks

When to choose

• you need a compact reasoning model trainable on tiny datasets without pre-training
• you are researching recurrent or hierarchical architectures for reasoning
• you have CUDA GPUs and want to reproduce the paper's ARC/Sudoku/maze results

When to avoid

• you need a production text-generation or chat model out of the box
• you lack NVIDIA GPUs or cannot build CUDA/FlashAttention extensions
• you want a plug-and-play library rather than research code

Facets

library · maturity active

deep-learning machine-learning llm-training deep-learning large-language-models artificial-intelligence machine-learning python hierarchical-reasoning-model recurrent-architecture reasoning brain-inspired-ai arc-benchmark research-code pytorch gpu linux

Member repositories

For agents

markdown · JSON · MCP: product_card(name="sapientinc/HRM")