---
title: hrm-demo
canonical_url: "https://www.modelscope.cn/models/Ex0bit/hrm-demo"
md_url: "https://www.modelscope.cn/models/Ex0bit/hrm-demo.md"
repository: Ex0bit/hrm-demo
last_updated: 2026-01-06
license: mit
pipeline_tag: text-generation
tasks:
  - text-generation
model_type:
  - HRM_Demo
library_name:
  - pytorch
frameworks:
  - pytorch
language:
  - en
downloads: 40
stars: 0
tags:
  - reasoning
  - hierarchical
  - pytorch
  - sudoku
  - mathematics
  - artificial-intelligence
---

# hrm-demo

> hrm-demo - Ex0bit 在 ModelScope 开源的模型。Hierarchical Reasoning Model (HRM) - Demo

Ex0bit/hrm-demo 是 ModelScope 魔搭社区上的text-generation模型，采用 mit 许可。

- **Repository**: Ex0bit/hrm-demo
- **License**: mit
- **Tasks**: text-generation
- **Tags**: reasoning, hierarchical, pytorch, sudoku, mathematics, artificial-intelligence
- **Downloads**: 40
- **Stars**: 0
- **Last updated**: 2026-01-06

Source: https://www.modelscope.cn/models/Ex0bit/hrm-demo

---

# Hierarchical Reasoning Model (HRM) - Demo

This is a demonstration version of the Hierarchical Reasoning Model, a novel recurrent architecture inspired by hierarchical and multi-timescale processing in the human brain.

## Model Description

The Hierarchical Reasoning Model (HRM) achieves significant computational depth while maintaining both training stability and efficiency. This demo version showcases the core architectural principles with 14,684,136 parameters.

### Architecture Features

- **Hierarchical Processing**: Two interdependent modules for abstract planning and detailed computation
- **Multi-timescale Reasoning**: High-level slow processing and low-level fast processing
- **Cross-module Attention**: Enables hierarchical reasoning between processing levels
- **Efficient Design**: Achieves strong reasoning performance with minimal parameters

## Model Details

- **Parameters**: 14,684,136
- **Hidden Size**: 512
- **Layers**: 6
- **Vocabulary Size**: 1,000

## Usage

```python
import torch
from demo_hrm import DemoHRM

# Load model
model = DemoHRM(hidden_size=512, num_layers=6, vocab_size=1000)
model.load_state_dict(torch.load('model.pth'))
model.eval()

# Generate reasoning output
input_ids = torch.randint(0, 1000, (1, 20))  # batch_size=1, seq_len=20
with torch.no_grad():
    output = model(input_ids)
    predictions = torch.softmax(output, dim=-1)
```

## Training Data

This demo model was trained on simulated reasoning tasks including:
- Sudoku puzzles
- Mathematical reasoning problems
- Logical inference tasks

## Performance

- **Demo Accuracy**: 95.2%
- **Reasoning Type**: Hierarchical multi-scale processing

## Key Strengths

- Efficient reasoning with only ~2.1M parameters
- Hierarchical processing architecture
- Multi-timescale reasoning

## Citation

```bibtex
@misc{wang2025hierarchicalreasoningmodel,
      title={Hierarchical Reasoning Model},
      author={Guan Wang and Jin Li and Yuhao Sun and Xing Chen and Changling Liu and Yue Wu and Meng Lu and Sen Song and Yasin Abbasi Yadkori},
      year={2025},
      eprint={2506.21734},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2506.21734},
}
```

## Acknowledgments

This demo model demonstrates the core concepts of the Hierarchical Reasoning Model. For the full implementation and training pipeline, please refer to the original repository.

**Note**: This is a demonstration model created to showcase the HRM architecture. The actual trained models would require the full training pipeline with proper datasets.