Core Technologies of Tencent Hunyuan AI Infra Made Open Source
12 hour ago / Read about 0 minute
Author:小编   

On February 4, 2026, the Tencent Hunyuan AI Infra team took a significant step by officially open-sourcing HPC-Ops, a production-grade, high-performance core operator library designed for large language model (LLM) inference. In practical, real-world applications, leveraging HPC-Ops led to a remarkable 30% surge in the queries per minute (QPM) for the Hunyuan model and a 17% increase in QPM for the DeepSeek model. When it comes to the performance of individual operators, the Attention operator witnessed a maximum performance boost of up to 2.22 times. The GroupGEMM operator saw its performance peak at a 1.88-fold increase, while the FusedMoE operator achieved a maximum improvement of 1.49 times.