Hunyuan-A13B, an open-source large language model built on a Mixture-of-Experts (MoE) architecture, has been detailed in a new technical report shared on Hugging Face. According to the paper, the model holds 80 billion total parameters but activates only 13 billion during inference, a design intended to balance capability, computational efficiency, and deployment cost.
The report states that the model was pretrained on a rigorously filtered corpus of 20 trillion tokens, with enhanced curation of STEM data aimed at improving factual reliability and reasoning. The authors say high-quality supervised fine-tuning and large-scale reinforcement learning were then applied to further boost overall performance.
A notable feature is a dual-mode Chain-of-Thought framework that adjusts reasoning depth to task complexity: a fast-thinking mode for routine queries and a slow-thinking mode for complex, multi-step problems. Evaluations cited in the report show competitive results across mathematics, science, programming, general language understanding, and agent tasks, often approaching the performance of much larger models.
The paper highlights high inference throughput, positioning Hunyuan-A13B as suitable for latency-sensitive applications. The model is being released to support open research and practical LLM deployment.