2026-09-23 Daily Report: Hunyuan-A13B Technical Report

Hunyuan-A13B Technical Report

Authors Tencent Hunyuan Team, Ao Liu, Botong Zhou, Can Xu, Chayse Zhou, ChenChen Zhang, Chengcheng Xu, Chenhao Wang, Decheng Wu, Dengpeng Wu, Dian Jiao, Dong Du, Dong Wang, Feng Zhang, Fengzong Lian, Guanghui Xu, Guanwei Zhang, Hai Wang, Haipeng Luo, Han Hu, Huilin Xu, Jiajia Wu, Jianchen Zhu, Jianfeng Yan, Jiaqi Zhu, Jihong Zhang, Jinbao Xue, Jun Xia, Junqiang Zheng, Kai Liu, Kai Zhang, Kai Zheng, Kejiao Li, Keyao Wang, Lan Jiang, Lixin Liu, Lulu Wu, Mengyuan Huang, Peijie Yu, Peiqi Wang, Qian Wang, Qianbiao Xiang, Qibin Liu, Qingfeng Sun, Richard Guo, Ruobing Xie, Saiyong Yang, Shaohua Chen, Shihui Hu, Shuai Li, Shuaipeng Li, Shuang Chen, Suncong Zheng, Tao Yang, Tian Zhang, Tinghao Yu, Weidong Han, Weijie Liu, Weijin Zhou, Weikang Wang, Wesleye Chen, Xiao Feng, Xiaoqin Ren, Xingwu Sun, Xiong Kuang, Xuemeng Huang, Xun Cao, Yanfeng Chen, Yang Du, Zhen Yang, Yangyu Tao, Yaping Deng, Yi Shen, Yigeng Hong, Yiqi Chen

Affiliations Tencent

Categories Method / Mixture-of-Experts / MoE architecture for large language model, Application / Foundation Model / Open-source large-scale LLM, Evaluation / Model Efficiency / Parameter count vs inference cost

License CC BY 4.0

Abstract Overview

Hunyuan-A13B is an open-source Mixture-of-Experts language model from Tencent featuring 80 billion total parameters while activating only 13 billion parameters per token. It is pretrained on a filtered 20T-token corpus that incorporates 250 billion tokens of curated STEM data, followed by context expansion to 256K tokens. The post-training pipeline integrates two distinct phases of supervised fine-tuning and reinforcement learning: one focused on complex reasoning and another covering broad real-world scenarios. The model also introduces a dual-mode Chain-of-Thought framework offering selectable fast-thinking and slow-thinking inference modes to balance response speed and reasoning depth.

Novelty

The technical report introduces a sparse 80B/13B fine-grained MoE architecture paired with a dual-mode Chain-of-Thought mechanism that allows explicit switching between short and deep reasoning at inference time. It also details a sequential post-training methodology separating verifiable reasoning optimization from multi-scenario general alignment and agent training.

Results

In base model evaluations, Hunyuan-A13B outperforms Qwen2.5-72B on nearly all reported metrics and surpasses Qwen3-A22B on 7 of 12 benchmarks despite using fewer active parameters. In post-trained slow-thinking evaluations, it achieves top scores among compared models on AIME2024 (87.3), BBH (89.1), ZebraLogic (84.7), and agent benchmarks including BFCL v3 (78.3), ComplexFuncBench (61.2), and C3-Bench (63.5). In long-context QA, it scores 73.9 on the 64K–128K segment of RULER, outperforming DeepSeek-R1 and Qwen3-A22B while delivering inference throughput up to 1981.99 tokens/s at batch size 32.

Key Points

Hunyuan-A13B adopts a fine-grained MoE architecture with 1 shared expert and 64 specialized experts (8 activated), utilizing 13B active parameters out of 80B total parameters.
The training recipe encompasses a 20T-token pretraining stage enriched with STEM data, context scaling up to 256K tokens, and a dual-stage post-training pipeline with user-controlled fast/slow Chain-of-Thought modes.
The model demonstrates competitive reasoning and agent performance against larger models while sustaining strong long-context stability and high inference throughput.

References

arXiv: https://arxiv.org/abs/2609.27284v1
Fugu-MT: https://fugumt.com/fugumt/paper_check/2609.27284v1
Hugging Face Papers: https://huggingface.co/papers/2609.27284