We introduce Hunyuan‑A13B, an open‑source large language model built on a Mixture‑of‑Experts (MoE) architecture. The model contains a total of 80 B parameters, but only about 13 B are activated during inference, striking a balance between capability, computational efficiency, and deployment cost.
Pre‑training is performed on a rigorously filtered 20 T‑token corpus with an emphasis on curated STEM data, which improves factual reliability and reasoning ability. High‑quality supervised fine‑tuning followed by large‑scale reinforcement learning further boosts overall performance.
Hunyuan‑A13B also proposes a dual‑mode Chain‑of‑Thought framework that adapts reasoning depth to task complexity: a fast‑thinking mode for routine queries and a slow‑thinking mode for complex, multi‑step problems, achieving both speed and accuracy.
Evaluations across mathematics, science, programming, general language understanding, and agent tasks show competitive results, often approaching those 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.
Review