How Tencent’s HunYuan Model Dominated All Major Video Retrieval Benchmarks

Today, Tencent announced that its "HunYuan" AI model has achieved a grand‑slam by securing first place on the five most authoritative cross‑modal video retrieval benchmarks: MSR‑VTT, MSVD, LSMDC, DiDeMo, and ActivityNet.

On the MSR‑VTT leaderboard, HunYuan raised text‑to‑video retrieval accuracy to 55%, leading the second‑place model by 1.7% and claiming the top industry spot.

The model spans multiple AI domains—including computer vision, natural language processing, multimodal content understanding, copy generation, and text‑to‑video synthesis—and is built on Tencent’s Taiji machine‑learning platform, leveraging GPU power for rapid algorithm iteration and model training.

Developed by Tencent Advertising’s Multimedia AI team, the "HunYuan_tvr" model introduces a novel hierarchical cross‑modal technique that decomposes video and text streams, performs similarity analysis, and extracts layered semantic relationships between them.

This "layer‑first, associate‑later, retrieve‑last" approach captures fine‑grained semantic information within each modality while effectively linking multimodal data, resulting in a substantial boost in retrieval precision.

The marked improvement in accuracy brings computer vision closer to human‑level video comprehension and signals a breakthrough in domestic multimodal research, offering long‑term value for both academic studies and industrial deployments.

HunYuan is already widely applied in Tencent’s advertising creation, retrieval, and recommendation scenarios, helping creators predict video‑content interest for specific audiences, enhancing creation efficiency, and increasing recommendation precision for a better user experience.

As image‑and‑video content continues to dominate online media, fine‑grained video understanding and multimodal feature fusion have become critical research priorities, prompting many AI companies to invest heavily in this field.

The five benchmark datasets—MSR‑VTT, MSVD, LSMDC, DiDeMo, and ActivityNet—are organized by leading institutions such as Microsoft, UC Berkeley, and King Abdullah University of Science and Technology, and serve as key arenas for showcasing AI model capabilities.