Hunyuan-Large is a large-scale Mixture-of-Experts (MoE) language model developed by Tencent. It features a total of 389 billion parameters, with 52 billion active parameters per token, designed to balance high-capacity performance with computational efficiency. This architecture allows the model to compete with leading dense models while significantly reducing the resource overhead required for inference.\n\nThe model supports an extensive context window of 256,000 tokens, enabling it to analyze long documents and maintain coherence across complex, multi-turn interactions. It was trained on a massive dataset exceeding 7 trillion tokens, which includes a substantial proportion of high-quality synthetic data produced through an iterative refinement process. This training methodology specifically targets improvements in logical reasoning, mathematical problem-solving, and professional coding tasks.\n\nHunyuan-Large is part of Tencent's broader ecosystem of open-weights models. Its technical design incorporates architectural optimizations such as Grouped Query Attention (GQA), Cross-Layer Attention (CLA), and Rotary Positional Embeddings (RoPE). The model uses a mixed routing strategy that combines shared and specialized experts to capture a diverse range of knowledge, facilitating its application across various multilingual and specialized domains.