ReasonFlux-F1-7B/README.md

library_name, license, base_model, tags, model-index

library_name	license	base_model	tags	model-index
transformers	other	deepseek-ai/DeepSeek-R1-Distill-Qwen-7B	llama-factory full generated_from_trainer	name results ReasonFlux-F1-7B

ReasonFlux: Hierarchical LLM Reasoning via Scaling Thought Templates

Revolutionary template-augmented reasoning paradigm enpowers a 32B model to outperform o1-mini and DeepSeek-R1 distilled models in reasoning tasks.

Task/Pass@1	ReasonFlux-F1-32B	ReasonFlux-Zero-32B	R1-Distill-32B	o1-mini	LIMO -32B	s1-32B
MATH500	96.0	91.2	94.3	90.0	90.6	93.0
AIME 2024	76.7	56.7	72.6	56.7	50.0	56.7
AIME 2025	53.3	37.2	46.67	50.8	37.2	49.3
GPQA-Diamond	67.2	61.2	62.1	60.0	65.2	59.6

ReasonFlux-F1-7B

ReasonFlux-F1-7B is our finetuned SOTA-level reasoning LLM by leveraging the template-augmented reasoning trajectories from our ReasonFlux-Zero.

• Github Repository: Gen-Verse/ReasonFlux
• Paper:ReasonFlux: Hierarchical LLM Reasoning via Scaling Thought Templates
• Dataset: Gen-Verse/ReasonFlux-F1-SFT

Evaluation

We present the evaluation results of our ReasonFlux-F1-32B on challenging reasoning tasks including AIME2024,AIM2025,MATH500 and GPQA-Diamond. To make a fair comparison, we report the results of the LLMs on our evaluation scripts in ReasonFlux-F1.

Model	AIME2024@pass1	AIME2025@pass1	MATH500@pass1	GPQA@pass1
QwQ-32B-Preview	46.7	37.2	90.6	65.2
LIMO-32B	56.3	44.5	94.8	58.1
s1-32B	56.7	49.3	93.0	59.6
OpenThinker-32B	66.0	53.3	94.8	60.1
R1-Distill-32B	70.0	46.7	92.0	59.6
ReasonFlux-Zero-32B	56.7	37.2	91.2	61.2
ReasonFlux-F1-32B	76.7	53.3	96.0	67.2

Quick start with VLLM

from vllm import LLM, SamplingParams
from transformers import AutoTokenizer

model_id = 'Gen-Verse/ReasonFlux-F1-7B'

model = LLM(
    model_id,
    tensor_parallel_size=8,
)
tokenizer = AutoTokenizer.from_pretrained(model_id)

sampling_params = SamplingParams(
    max_tokens=32768,
)
# 2022 AIME I Problems/Problem 15
question = """Let \(x, y\), and \(z\) be positive real numbers satisfying the system of equations:
\[
\begin{array}{c}
\sqrt{2 x-x y}+\sqrt{2 y-x y}=1 \\
\sqrt{2 y-y z}+\sqrt{2 z-y z}=\sqrt{2} \\
\sqrt{2 z-z x}+\sqrt{2 x-z x}=\sqrt{3} .
\end{array}
\]
Then \(\left[(1-x)(1-y)(1-z)\right]^{2}\) can be written as \(\frac{m}{n}\), where \(m\) and \(n\) are relatively prime positive integers. Find \(m+n\)."""
ds_prompt="<｜User｜>\n" + question + "<｜Assistant｜>\n"
output = model.generate(ds_prompt, sampling_params=sampling_params)
print(output[0].outputs[0].text)

Citation

@article{yang2025reasonflux,
  title={ReasonFlux: Hierarchical LLM Reasoning via Scaling Thought Templates},
  author={Yang, Ling and Yu, Zhaochen and Cui, Bin and Wang, Mengdi},
  journal={arXiv preprint arXiv:2502.06772},
  year={2025}
}