metadata

license: apache-2.0
datasets:
  - Multiverse4FM/Multiverse-1K-mixed
  - Multiverse4FM/Multiverse-1K
  - simplescaling/s1K-1.1
base_model:
  - Qwen/Qwen2.5-32B-Instruct
pipeline_tag: text-generation
library_name: transformers

Model Summary

Multiverse-32B, built on Multiverse, is the first open-source, non-AR model to achieve scores of 54% and 46% on AIME 2024 & 2025.

Webpage: Multiverse
Paper: https://arxiv.org/abs/2506.09991

Use

The model usage is documented here.

Evaluation

Model	AIME24	AIME25	MATH500	GPQA-Diamond
s1-32B	35.4	25.8	88.6	48.0
s1.1-32B	52.9	41.7	93.4	62.6
Qwen2.5-32B-Instruct	15.8	10.4	80.4	47.0
Autoregressive-32B	54.6	45.0	92.8	61.6
Multiverse-32B-zero	52.1	44.2	92.4	63.6
Multiverse-32B	53.8	45.8	91.8	60.7

Acknowledge

Thanks to the amazing s1 team for their s1.1 dataset as base data, and the Qwen team for their Qwen-2.5-32B-Instruct as base model.

Citation Information

@misc{yang2025multiverselanguagemodelssecretly,
      title={Multiverse: Your Language Models Secretly Decide How to Parallelize and Merge Generation}, 
      author={Xinyu Yang and Yuwei An and Hongyi Liu and Tianqi Chen and Beidi Chen},
      year={2025},
      eprint={2506.09991},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2506.09991}, 
}