Open-o3

Blog | Website | Feedback

Tip

This is a framework of reasoning, not a model. You can use the framework on any LLM.

Open-o3 is the open-source version of OpenAI's o3. This project aims to provide a powerful, open, and accessible language model for researchers and developers, advancing the field of artificial intelligence. Open-o3 is committed to offering flexible and customizable interfaces to support various reasoning tasks such as numerical calculations, programming, logic reasoning, etc. Through this open-source project, we hope to foster innovation and application of AI technology.

Project Description

Open-o3 is an efficient and scalable language model, which is an open-source reproduction of OpenAI's o3 model. Our goal is to provide developers and researchers with a reliable and easy-to-use tool to help them implement and explore smarter and more complex natural language processing applications.

Features

• Systematic Resampling: Implements OpenAI's o3 approach to exponentially improve accuracy through iterative reasoning.
• Model-Agnostic: Compatible with any LLM, allowing you to leverage your preferred model.
• Result Tracking: Save and analyze reasoning paths to improve performance.
• Customizable Prompts: Tailor the system prompt to your specific use cases and domains.
• Extensible Architecture: Easily build upon the core framework for specialized applications.

Quick Start

Basic Usage

from open_o3 import o3

system_prompt = "You are a large reasoning model Open-o3 developed by Lumen Lab."
reasoner = o3(system_prompt=system_prompt, enable_tda=True, save_results=True)
reasoner.run_example()

Advanced Configuration

from open_o3 import o3

# Custom configuration
reasoner = o3(
system_prompt="You are specialized in solving mathematical problems.",
model="deepseek-r1",
max_iterations=5,
temperature=0.7,
enable_tda=True,
save_results=True,
result_path="./reasoning_logs/"
)

# Solve a specific problem
result = reasoner.solve("What is the integral of x²?")
print(result)

Parameters

Parameter	Type	Default	Description
system_prompt	str	Required	Initial instructions for the model
model	str	"deepseek-r1"	The LLM to use
enable_tda	bool	False	Whether to use TDA
max_iterations	int	3	Maximum reasoning iterations
temperature	float	0.8	Sampling temperature for the model
save_results	bool	False	Whether to save reasoning paths
result_path	str	". /results/"	Path to save reasoning logs

Examples

Mathematical Reasoning

result = reasoner.solve("Solve for x: 2x² + 5x - 3 = 0")

Code Generation

code = reasoner.solve("Write a function to check if a string is a palindrome in Python")

Logical Reasoning

analysis = reasoner.solve("If all A are B, and some B are C, what can we conclude about A and C?")

Usage Example

system_prompt = "You are a large reasoning model Open-o3 developed by Lumen Lab."
reasoner = o3(system_prompt=system_prompt, enable_tda=True, save_results=True)
reasoner.run_example()

Development and Contribution

We welcome contributions in any form, whether it's reporting bugs or submitting new features. To contribute, you can:

1. Create a new issue to report a problem or suggest a new feature.
2. Fork the repository and develop new features or fix bugs on your own branch.
3. Submit a pull request.

Roadmap

• Multi-model ensemble reasoning
• Fine-tuning support for specific domains
• Performance benchmarks against other reasoning frameworks
• Community templates for common reasoning tasks

Citation

If you use Open-o3 in your research, please cite:

@software{Open-o3,
author = {Lumen Lab},
title = {Open-o3: A Framework for Exponentially Improving LLM Accuracy},
url = {https://github.com/ChihayaYuka/Open-o3},
year = {2025},
}

License

This project is licensed under the MIT License, allowing you to freely use, modify, and distribute the code. For more details, please refer to the LICENSE file.

Contact Information

If you have any questions, suggestions, or feedback, feel free to contact us via the following methods:

• Email: yuka@lumenlab.cc
• GitHub Issues: https://github.com/ChihayaYuka/Open-o3