
Academic lab · Mountain View, United States
OpenThoughts
A research collective curating fully open reasoning datasets and recipes, from the OpenThoughts3 data recipe to the OpenThinker models.
Mountain View, United States
- Founded
- 2025
- Type
- Academic lab
- Based in
- Mountain View, United States
- Key papers
- 1
- Featured projects
- 7
About
OpenThoughts is a collaboration of researchers and engineers from universities including Stanford, UC Berkeley, UT Austin, NYU, the University of Washington and TUM, together with LAION. It launched publicly in January 2025 with the aim of producing open post-training data for reasoning models.
The project publishes datasets, training code and models together. Its OpenThoughts3 recipe, described in the paper "OpenThoughts: Data Recipes for Reasoning Models", was derived from more than 1,000 controlled experiments and produced the 1.2 million example OpenThoughts3 dataset and the OpenThinker3-7B model.
Compute and support come from academic centers such as the Jülich Supercomputing Centre, TACC and TU Dresden, as well as industry partners. From late 2025 the group extended its work to agentic tasks under the OpenThoughts-Agent project.
Mission and ambitionsOpenThoughts aims to make the data side of reasoning models as open as the weights, so that the community can study which data choices matter. It treats systematic, reproducible ablation of data pipelines as its main method.
Known for
- 01
OpenThoughts3 reasoning dataset
- 02
OpenThinker open reasoning models
- 03
Large-scale data-recipe ablations
- 04
OpenThoughts-Agent project
Milestones
- 2025
Project launch
The OpenThoughts project was announced with the goal of open reasoning data.
- 2025
OpenThinker-32B
Released an open-data 32B reasoning model, described by the project as scaling up open reasoning.
- 2025
OpenThinker2
Published a follow-up model family reported to outperform DeepSeek-R1-Distill-32B on its benchmarks.
- 2025
OpenThoughts3
Released the 1.2M example dataset and a recipe based on over 1,000 experiments, with OpenThinker3-7B.
- 2025
OpenThoughts-Agent
Launched a project on data recipes and benchmarks for terminal and agentic tasks, followed by the OpenThoughts-TBLite benchmark in 2026.
Open source
Fully open data curation for reasoning models
Data recipes and infrastructure for training agents
Benchmark for terminal agents
- OpenThoughts-114kdataset
Reasoning dataset
- OpenThoughts3-1.2Mdataset
1.2M-example reasoning dataset
- OpenThinker3-7Bmodel
7B reasoning model
- OpenThinker-Agent-v1model
Terminal-agent model
Research papers
Notable releases
Openness
Apache-2.0 code, with datasets and OpenThinker model weights released on Hugging Face.
People
- Etash GuhaLead author, OpenThoughts
First author of the OpenThoughts data recipes paper.
- Ludwig SchmidtAssistant professor, Stanford; LAION member
Last author of the OpenThoughts paper.
- Alexandros G. DimakisProfessor, UC Berkeley; Bespoke Labs co-founder
Co-author of OpenThoughts and OpenThoughts-Agent.
- Ryan MartenCo-author
Second author of the OpenThoughts paper.
Listed from the lab's own pages and public profiles.
Contact and community
Via website.
Sources
Compiled from the lab's public pages, Hugging Face and GitHub. Last checked 2026-11-17. Something wrong or out of date? Tell us on Discord.



