togethercomputer/M1-3B

This is the model is trained using paper, M1: Towards Scalable Test-Time Compute with Mamba Reasoning Models.

| Model | AIME 2025 | AIME 2024 | MATH 500 | AMC 2023 | OlympiadBench | |-----------------------------------|---------------|---------------|--------------|--------------|-------------------| | Qwen2.5-Math-7B-Instruct (Transformer) | – | 13.3 | 79.8 | 50.6 | 40.7 | | rStar-Math-7B (Transformer) | – | 26.7 | 78.4 | 47.5 | 47.1 | | Eurus-2-7B-PRIME (Transformer) | – | 26.7 | 79.2 | 57.8 | 42.1 | | Qwen2.5-7B-SimpleRL (Transformer) | – | 26.7 | 82.4 | 62.5 | 43.3 | | DeepSeek-R1-Distill-Qwen-1.5B (Transformer) | 23.0 | 28.8 | 82.8 | 62.9 | 43.3 | | M1-3B (Mamba Hybrid Models) | 23.5 | 28.5 | 84.0 | 62.8 | 47.3 |

Code: https://github.com/jxiw/M1

@article{wang2025m1scalabletesttimecompute,
title={M1: Towards Scalable Test-Time Compute with Mamba Reasoning Models},
author={Junxiong Wang and Wen-Ding Li and Daniele Paliotta and Daniel Ritter and Alexander M. Rush and Tri Dao},
journal={arXiv preprint arXiv:2504.10449},
year={2025},
url={https://arxiv.org/abs/2504.10449},
}