Open-O1 is a non-profit open-source project whose goal is to match the capabilities of the OpenAI O1 model and provide the community with an advanced open-source alternative. The project builds an SFT dataset for chain-of-thought (CoT) activation and uses this data to train LLaMA and Qwen models, thereby giving smaller models stronger long reasoning and problem-solving capabilities.
Main capabilities and outputs The project has released two models: OpenO1-LLaMA-8B-v0.1 and OpenO1-Qwen-7B-v0.1, available on HuggingFace; it has also open-sourced the OpenO1 SFT data for CoT activation for the community to download and use. According to the official description, OpenO1 mainly exhibits O1-like reasoning characteristics and a broad search-style thinking ability. Official benchmark tests (zero-shot setting) show: GSM8K score 85.82, MATH score 52.88, MMLU score 70.45, Hellaswag score 67.77, ARC-C score 86.52, BBH score 58.43; on benchmarks such as MATH, MMLU, ARC-C, and BBH, it outperforms llama3.1-8b-instruct (among which the official MATH score of 51.9 for llama3.1-8b-instruct comes from the CoT setting, and the reproduced zero-shot score is 47.42).
Usage For local deployment, the official steps are as follows: 1) clone the GitHub repository; 2) install the dependencies listed in Deployment/requirements.txt; 3) run the scripts provided by the repository. The optional models are OpenO1-Qwen-7B-v0.1 or OpenO1-LLaMA-8B-v0.1. Open O1's chat template follows the LLaMA3.1 template, and specific details can be found in chat-templates. Official reminder: models deployed on HuggingFace may be quantized, which may cause some performance degradation; to experience the original performance, it is recommended to deploy on your own infrastructure.
Applicable scenarios and development status Open O1 is currently in an early development stage. The official statement says it has not yet fully reached O1-level capabilities, but it has taken an important step. The first version demonstrated performance improvements on LLaMA and Qwen, with a style close to O1. The project welcomes the community to join and will continue to open-source all code and materials from the development process. Related questionsand feedback can be submitted through GitHub Issues; it is recommended to first check the FAQ and past issues, and use the dedicated template. The project is non-profit.
Limitations and information not yet provided The official statement clearly indicates that the current model has not yet reached O1 level; the HuggingFace deployment version may suffer performance degradation due to quantization. Apart from the above, information on the complete composition of the training data, hyperparameter details (the page is marked "Summarizing and updating..."), fees and pricing, commercial licensing, etc., is not yet provided.
