Hunyuan-A13B Technical Report
Tencent Hunyuan Team, Ao Liu, Botong Zhou, Can Xu, Chayse Zhou, ChenChen Zhang, Chengcheng Xu, Chenhao Wang and 67 others
Tencent Hunyuan
Published September 23, 2026 · on the feed September 24, 2026 · ▲ 14 · arXiv 2609.27284
The abstract
We present Hunyuan-A13B, an open-source large language model based on a Mixture-of-Experts architecture. It contains 80 billion total parameters but activates only 13 billion during inference, balancing model capability, computational efficiency, and deployment cost. The model is pretrained on a rigorously filtered 20T-token corpus with enhanced STEM data curation, improving factual reliability and reasoning ability. High-quality supervised fine-tuning and large-scale reinforcement learning further enhance its overall performance. Hunyuan-A13B also introduces a dual-mode Chain-of-Thought framework that adapts reasoning depth to task complexity: fast thinking for routine queries and slow thinking for complex, multi-step problems. Evaluations show competitive performance across mathematics, science, programming, general language understanding, and agent tasks, often approaching that of much larger models. Its high inference throughput makes it suitable for latency-sensitive applications. We release Hunyuan-A13B to support open research and practical LLM deployment.
The Brief
1:49A short podcast summary of the paper. Written and voiced by AI.
Transcript
Hunyuan-A13B Technical Report. By Tencent Hunyuan Team and seventy-four others, from Tencent Hunyuan. Published on the twenty-third of September. No code repository is linked to it.
Hunyuan-A13B is an open-source large language model, released to support open research and practical deployment.
It is built on a Mixture-of-Experts architecture with eighty billion total parameters, of which only thirteen billion are active during inference. That split balances model capability, computational efficiency and deployment cost. The model also introduces a dual-mode Chain-of-Thought framework that adapts reasoning depth to task complexity: fast thinking for routine queries, slow thinking for complex, multi-step problems. The two modes let one model spend little on easy requests and more where the task demands it, rather than applying one depth of reasoning to everything.
Pretraining used a rigorously filtered corpus of twenty trillion tokens, with enhanced curation of STEM data, aimed at factual reliability and reasoning ability. High-quality supervised fine-tuning and large-scale reinforcement learning then further enhanced overall performance. Evaluations show competitive performance across mathematics, science, programming, general language understanding and agent tasks, often approaching that of much larger models. The recipe, in short, runs from careful data, through supervised fine-tuning, to reinforcement learning at scale.
It is aimed at open research and at practical deployment of language models. Its high inference throughput makes it suitable for applications that are sensitive to latency.
The central trade is in the name. Thirteen billion parameters are active out of eighty billion in total, so the model carries the larger capacity while activating only a fraction of it. A second lever sits on top: deciding per query whether to think fast or slow.
Paper Quest · 7 chapters · about 40 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Strong models are expensive to run. That cost is the wall here.
You want an open language model strong enough to approach the best systems. But leading models demand huge compute at inference, with high latency and hardware costs that put them out of reach for many.
The road ahead
- 1 · The problem ~7 min The strongest language models cost so much to run that many people cannot afford them, because every part of a dense model works on every input.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Hold many expert networks but run only a few per input, keep one shared expert always on, and let one model think fast or slow.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Build a sparse expert model, pretrain it on carefully filtered data, then fine-tune it on high-quality examples and improve it with large-scale reinforcement learning.
- 6 · The experiment ~6 min You’ll guess: With only 13 billion active parameters, how did Hunyuan-A13B perform against much larger models across maths, science, programming and agent tasks?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.