LightTransfer

LightTransfer: Your Long-Context LLM is Secretly a Hybrid Model with Effortless Adaptation

💻Github Repo • 🤔Reporting Issues • 📜Paper Link

Model Card for Model ID

Model Details:

Base Model: Qwen/Qwen2.5-32B-Instruct
datasets: RUC-AIBOX/long_form_thought_data_5k
Training Framework: Supervised Fine-tuning
Parameters: 32B
Special Features:
- Replace 50% full attention layers with streaming attention

Model Details

QwQ-LightTransfer is a 32B-parameter model built on Qwen/Qwen2.5-32B-Instruct and fine-tuned via SFT on RUC-AIBOX/long_form_thought_data_5k.

By replacing 50% of the model’s full attention layers with streaming attention,specifically layers [5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 26, 27, 28, 30, 31, 32, 33, 35, 37, 38, 43, 51], it substantially reduces memory costs.
QwQ-LightTransfer scores 53.3% on the advanced math benchmark AIME24, demonstrating its strong o1-like long reasoning capabilities.

Performance Evaluation

We have evaluated QwQ-LightTransfer on several long reasoning generation benchmarks. Some of the evaluation results are shown in the table below.

Method	Math-OAI	AIME24	AIME25	GSM8K
o1-preview	85.5	44.6	-	-
OwO-STILL	90.2	46.7	33.3	95.6
LongGen	78.2	16.7	-	95.4
LightTransfer	90.7	53.3	40.0	95.5

Import from Transformers

To load the QwQ-LightTransfer model using Transformers, use the following code:

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
model_name = 'QwQ-32B-LightTransfer'
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name,torch_dtype=torch.bfloat16,trust_remote_code=True,device_map='auto')

text = "Hi, I'm QwQ-32B-LightTransfer."
inputs = tokenizer(text, return_tensors='pt').to(model.device)
with torch.no_grad():
    outputs = model.generate(inputs['input_ids'],max_gen_len=32000)

print(tokenizer.decode(outputs[0]))

Citation

@misc{zhang2025lighttransferlongcontextllmsecretly,
      title={LightTransfer: Your Long-Context LLM is Secretly a Hybrid Model with Effortless Adaptation}, 
      author={Xuan Zhang and Fengzhuo Zhang and Cunxiao Du and Chao Du and Tianyu Pang and Wei Gao and Min Lin},
      year={2025},
      eprint={2410.13846},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2410.13846}, 
}