[!NOTE] We have released a paper for OpenThoughts! See our paper here.
OpenThinker-32B
This model is a fine-tuned version of Qwen/Qwen2.5-32B-Instruct on the OpenThoughts-114k dataset.
The dataset is derived by distilling DeepSeek-R1 using the data pipeline available on github. More info about the dataset can be found on the dataset card at OpenThoughts-114k dataset.
The numbers reported in the table below are evaluated with our open-source tool Evalchemy.
| Model Name | Dataset Size | AIME24 I/II | AIME25 I | MATH500 | GPQA Diamond | LCBv2 |
|---|---|---|---|---|---|---|
| LIMO-32B | 0.8k | 56.7 | 49.3 | 86.6 | 58.1 | 60.0 |
| s1-32B | 1k | 36.0 | 25.3 | 84.8 | 50.5 | 40.9 |
| s1.1-32B | 1k | 64.7 | 49.3 | 89.0 | 60.1 | 65.5 |
| DeepSeek-R1-Distill-Qwen-32B | 800k (closed) | 76.7 | 55.9 | 89.4 | 57.6 | 71.2 |
| OpenThinker-32B | 114k | 66.0 | 53.3 | 90.6 | 61.6 | 68.9 |
We are fully open-source. Our model weights, datasets, data generation code, evaluation code, and training code are all publicly available.
| Open Weights | Open Data | Open Code | |
|---|---|---|---|
| OpenThinker-32B | ✅ | ✅ | ✅ |
| DeepSeek-R1-Distill-Qwen-32B | ✅ | ❌ | ❌ |
| OpenAI/Gemini | ❌ | ❌ | ❌ |
Intended uses & limitations
Apache 2.0 License
Training procedure
We finetune Qwen2.5-32B-Instruct on OpenThoughts-114k for 3 epochs with a 16k context length using LlamaFactory. Our full training configuration is provided in our repository. Training the 32B model on OpenThoughts-114k was done on AWS SageMaker with 8xH100 P5 nodes. On 4 nodes, this took around 90 hours. Meanwhile, for training on OpenThoughts-Unverified-173k, we used 96 nodes of 4xA100 (64 GB per GPU), training took 30 hours, spending 11,520 A100 hours on the Leonardo Supercomputer.
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 1e-05
- train_batch_size: 1
- eval_batch_size: 8
- seed: 42
- distributed_type: multi-GPU
- num_devices: 32
- gradient_accumulation_steps: 3
- total_train_batch_size: 96
- total_eval_batch_size: 256
- optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
- lr_scheduler_type: cosine
- lr_scheduler_warmup_ratio: 0.1
- num_epochs: 3.0
Framework versions
- Transformers 4.46.1
- Pytorch 2.3.0
- Datasets 3.1.0
- Tokenizers 0.20.3
More info can be found in our repository: https://github.com/open-thoughts/open-thoughts.
Links
- 📝 OpenThoughts Paper
- 📊 Open Thoughts Launch Blog Post
- 📊 Open Thoughts Measuring Reasoning with Evalchmey Blog Post
- 📊 Open Thoughts OpenThinker-32B Post
- 💻 Open Thoughts GitHub Repository
- 🧠 OpenThoughts-114k dataset
- 🧠 OpenThoughts-Unverified-173k dataset
- 🤖 OpenThinker-7B model
- 🤖 OpenThinker-7B-Unverfied model
- 🤖 OpenThinker-32B model - this model
- 🤖 OpenThinker-32B-Unverified model
Citation
@misc{guha2025openthoughtsdatarecipesreasoning,
title={OpenThoughts: Data Recipes for Reasoning Models},
author={Etash Guha and Ryan Marten and Sedrick Keh and Negin Raoof and Georgios Smyrnis and Hritik Bansal and Marianna Nezhurina and Jean Mercat and Trung Vu and Zayne Sprague and Ashima Suvarna and Benjamin Feuer and Liangyu Chen and Zaid Khan and Eric Frankel and Sachin Grover and Caroline Choi and Niklas Muennighoff and Shiye Su and Wanjia Zhao and John Yang and Shreyas Pimpalgaonkar and Kartik Sharma and Charlie Cheng-Jie Ji and Yichuan Deng and Sarah Pratt and Vivek Ramanujan and Jon Saad-Falcon and Jeffrey Li and Achal Dave and Alon Albalak and Kushal Arora and Blake Wulfe and Chinmay Hegde and Greg Durrett and Sewoong Oh and Mohit Bansal and Saadia Gabriel and Aditya Grover and Kai-Wei Chang and Vaishaal Shankar and Aaron Gokaslan and Mike A. Merrill and Tatsunori Hashimoto and Yejin Choi and Jenia Jitsev and Reinhard Heckel and Maheswaran Sathiamoorthy and Alexandros G. Dimakis and Ludwig Schmidt},
year={2025},
eprint={2506.04178},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2506.04178},
}