metadata

pipeline_tag: text-generation
inference: true
license: apache-2.0
datasets:
  - simplescaling/s1K

We recommend using our successor s1.1 with better performance

Model Summary

s1 is a reasoning model finetuned from Qwen2.5-32B-Instruct on just 1,000 examples. It matches o1-preview & exhibits test-time scaling via budget forcing.

Repository: simplescaling/s1
Paper: https://arxiv.org/abs/2501.19393

Use

The model usage is documented here.

Evaluation

Metric	s1-32B	s1.1-32B	o1-preview	o1	DeepSeek-R1	DeepSeek-R1-Distill-Qwen-32B
# examples	1K	1K	?	?	>800K	800K
AIME2024	56.7	56.7	40.0	74.4	79.8	72.6
AIME2025 I	26.7	60.0	37.5	?	65.0	46.1
MATH500	93.0	95.4	81.4	94.8	97.3	94.3
GPQA-Diamond	59.6	63.6	75.2	77.3	71.5	62.1

Note that s1-32B and s1.1-32B use budget forcing in this table; specifically ignoring end-of-thinking and appending "Wait" once or twice.

Citation

@misc{muennighoff2025s1simpletesttimescaling,
      title={s1: Simple test-time scaling}, 
      author={Niklas Muennighoff and Zitong Yang and Weijia Shi and Xiang Lisa Li and Li Fei-Fei and Hannaneh Hajishirzi and Luke Zettlemoyer and Percy Liang and Emmanuel Candès and Tatsunori Hashimoto},
      year={2025},
      eprint={2501.19393},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2501.19393}, 
}