# Quantization Hurts Reasoning? An Empirical Study on Quantized Reasoning Models

**Authors:** Ruikang Liu, Yuxuan Sun, Manyi Zhang, Haoli Bai, Xianzhi Yu, Tiezheng Yu, Chun Yuan, Lu Hou  
**arXiv:** [2504.04823](https://arxiv.org/abs/2504.04823)  
**Date:** 2025-04-07  
**Summary page:** https://www.summarizepaper.com/en/paper/2504.04823v2/

## Key points

- Authors conducted a study on the impact of quantization on reasoning language models
- Quantization aims to reduce inference cost but its effects on reasoning models have not been thoroughly explored
- Evaluation of various model families across a wide parameter range from 1.5B to 70B parameters
- Lossless quantization possible with certain methods, but lower bit-widths introduce accuracy risks
- Model size, origin, and task difficulty are crucial factors influencing performance
- Quantized models do not show increased output lengths as expected
- Scaling model sizes or reasoning steps can enhance performance

## Summary

, , , , 

In their paper titled "Quantization Hurts Reasoning? An Empirical Study on Quantized Reasoning Models," authors Ruikang Liu, Yuxuan Sun, Manyi Zhang, Haoli Bai, Xianzhi Yu, Tiezheng Yu, Chun Yuan, and Lu Hou delve into the impact of quantization on reasoning language models. These models have shown impressive performance in complex tasks but come with increased inference overhead due to their extended chain-of-thought reasoning process. While quantization has been widely used to reduce the inference cost of large language models, its effects on reasoning models have not been thoroughly explored. The study conducted by the authors is the first systematic investigation into quantized reasoning models. They evaluate various open-sourced model families such as DeepSeek-R1-Distilled Qwen and LLaMA across a wide parameter range from 1.5B to 70B parameters, including QwQ-32B and Qwen3-8B. The evaluation covers weight, KV cache, and activation quantization using cutting-edge algorithms at different bit-widths. The findings of the study reveal that while lossless quantization can be achieved with W8A8 or W4A16 quantization methods, lower bit-widths introduce significant accuracy risks. The authors also identify model size, model origin, and task difficulty as crucial factors influencing performance. Surprisingly, quantized models do not exhibit increased output lengths as expected. Additionally, strategically scaling model sizes or reasoning steps can effectively enhance performance. To facilitate further research in this area,<kgd> all quantized models and codes used in the study are openly available on GitHub at https://github.com/ruikangliu/Quantized-Reasoning-Models.