# Nanbeige4-3B Technical Report: Exploring the Frontier of Small Language Models

**Authors:** Chen Yang, Guangyue Peng, Jiaying Zhu, Ran Le, Ruixiang Feng, Tao Zhang, Wei Ruan, Xiaoqi Liu, Xiaoxue Cheng, Xiyun Xu, Yang Song, Yanzipeng Gao, Yiming Jia, Yun Xing, Yuntao Wen, Zekai Wang, Zhenwei An, Zhicong Sun, Zongchao Chen  
**arXiv:** [2512.06266](https://arxiv.org/abs/2512.06266)  
**Date:** 2025-12-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2512.06266v1/

## Key points

- Nanbeige4-3B is a family of high-performing language models pretrained on 23T tokens and finetuned on over 30 million diverse instructions.
- The model incorporates a Fine-Grained Warmup-Stable-Decay (FG-WSD) training scheduler in pre-training to extend the scaling law for small language models, boosting performance through progressive refinement of data mixtures.
- Post-training includes improving SFT data quality through deliberative generation refinement and chain-of-thought reconstruction, leading to gains on complex tasks.
- Dual Preference Distillation (DPD) is used to distill Nanbeige4-3B's flagship reasoning model, resulting in further performance improvements.
- A multi-stage reinforcement learning phase enhances reasoning and human alignment abilities using verifiable rewards and preference modeling.

## Summary

, , , , 

We introduce Nanbeige4-3B, a family of high-performing language models that have been pretrained on 23T tokens and finetuned on over 30 million diverse instructions. Our model extends the boundary of the scaling law for small language models by incorporating a Fine-Grained Warmup-Stable-Decay (FG-WSD) training scheduler in pre-training. This progressive refinement of data mixtures across stages boosts model performance. In post-training, we improve the quality of SFT data through a joint mechanism that integrates deliberative generation refinement and chain-of-thought reconstruction, resulting in substantial gains on complex tasks. Following SFT, our flagship reasoning model is used to distill Nanbeige4-3B through Dual Preference Distillation (DPD), leading to further performance improvements. A multi-stage reinforcement learning phase leverages verifiable rewards and preference modeling to enhance abilities in both reasoning and human alignment. Our extensive evaluations demonstrate that Nanbeige4-3B outperforms models of comparable parameter scale and rivals much larger models across various benchmarks. The post-training pipeline includes a Cold Start Supervised Fine-tuning stage aimed at establishing a robust foundation for reasoning by focusing on high-quality reasoning data such as mathematical, code, and subject-area problem-solving tasks. Scaling SFT Instructions from hundreds of thousands to tens of millions of examples continues to produce substantial improvements on challenging reasoning benchmarks without early saturation. Overall Supervised Fine-Tuning further enhances the model's general abilities and task diversity by combining general conversation and writing data with agent-style interaction data, harder reasoning tasks, and code-related tasks. Deliberative generation refinement combined with chain completion is employed to improve the model's output quality on complex tasks. In conclusion, Nanbeige4-3B demonstrates superior performance compared to models of similar size across a wide range of benchmarks. Model checkpoints are available at https://huggingface.co/Nanbeige for further exploration and evaluation purposes.