Open Thoughts is a fully open-source reasoning model data curation project, jointly led by Bespoke Labs and the DataComp community, with participants including researchers and engineers from institutions such as Stanford University, UC Berkeley, the University of Washington, UT Austin, the Juelich Supercomputing Center, LAION, UCLA, UNC Chapel Hill, and the Toyota Research Institute. The project's primary goal is to curate high-quality reasoning datasets and train small state-of-the-art reasoning models that surpass DeepSeek-R1-Distill-Qwen-32B and DeepSeek-R1-Distill-Qwen-7B on math and code reasoning benchmarks.
The project has released multiple datasets and model results. In June 2025, the OpenThoughts3-1.2M dataset was released. This dataset contains 850,000 math problems, 250,000 code problems, and 100,000 science problems. The reasoning traces were generated using QwQ-32B, and it is the result of more than 1,000 experiments used to test various design choices in data curation. In April 2025, the OpenThoughts2-1M dataset was released, composed of OpenThoughts-114k, OpenR1-Math, and newly generated math and code reasoning data, generated by sampling from the best-performing methods after ablation experiments on 26 different problem generation methods. The earlier OpenThoughts-114k and OpenThinker-7B were released in January 2025, followed by models such as OpenThinker-32B. Related models have received 400,000 downloads on Ollama.
The project is fully open source, with model weights, datasets, data generation code, evaluation code, and training code all publicly available. According to public information, the training and evaluation code is marked as "coming soon." The data generation process involves setting the DeepSeek API key and setting HF_ORG to the organization ID; if pushing to a private repository, HF_PRIVATE=true must be set. The reasoning benchmark has been added to the open-source tool Evalchemy and compared with publicly reported scores. To reduce evaluation precision fluctuations, the project uses different random seeds for multiple evaluations and calculates the average score.
In terms of applicable scenarios, the project is mainly aimed at researchers and developers who need high-quality open-source reasoning datasets, and can be used to train math, code, and science reasoning models. The open-source community has derived multiple works based on OpenThoughts. For example, the training of models such as Light-R1-SFT, Traceback-12B, and Sky-T1 used related datasets. On Hugging Face, more than 190 public models have used OpenThoughts-114k for training, and more than 100 public models have used Bespoke-Stratos-17k for training. Ollama also provides quantized versions of OpenThinker-7B and OpenThinker-32B for running on local laptops. In terms of limitations, official materials do not provide specific information such as fees, pricing, platform restrictions, or detailed usage steps, and none is currently provided.reasoning models. The open-source community has derived multiple works based on OpenThoughts. For example, the training of models such as Light-R1-SFT, Traceback-12B, and Sky-T1 used related datasets. On Hugging Face, more than 190 public models have used OpenThoughts-114k for training, and more than 100 public models have used Bespoke-Stratos-17k for training. Ollama also provides quantized versions of OpenThinker-7B and OpenThinker-32B for running on local laptops. In terms of limitations, official materials do not provide specific information such as fees, pricing, platform restrictions, or detailed usage steps, and none is currently provided.
