Hunyuan-Large-Vision: Multimodal Understanding Model in Tencent's Hunyuan Vision Family 75

小编 · 2025-08-13

Tencent's Hunyuan-Large-Vision is the multimodal understanding model in the Hunyuan vision family. It uses an MoE architecture made up of a native-resolution Hunyuan ViT vision encoder with billions of parameters, an MLP connector, and an MoE language model with 389B parameters and 52B activated. It supports arbitrary-resolution images, video, and 3D spatial inputs, with strengthened multilingual understanding, suited for camera-based problem solving, video calls, video understanding, and creative copywriting.

Hunyuan-Large-Vision: Multimodal Understanding Model in Tencent's Hunyuan Vision Family

Tencent's Hunyuan-Large-Vision is a multimodal understanding model within the Hunyuan vision model family. It adopts a MoE architecture, consisting of a billion-parameter native-resolution Hunyuan ViT visual encoder, an MLP connector module, and an MoE language model with 389B parameters and 52B active parameters. It supports arbitrary-resolution images, video, and 3D spatial inputs, with enhanced multilingual scene understanding capabilities. It can be used in scenarios such as photo-based problem solving, video calls, video understanding, and copywriting creation.

📚 Model Architecture

Visual Encoder: A billion-parameter native-resolution Hunyuan ViT visual encoder, optimized for multimodal tasks. It supports native-resolution input and can accurately capture visual information from images and videos of various sizes.
MLP Connector Module: Designed with an adaptive downsampling mechanism, it efficiently compresses visual features and connects the visual encoder with the language model.
MoE Language Model: Contains 389B parameters and 52B active parameters, with strong multilingual understanding and reasoning capabilities.

📚 Performance Parameters

It scored 1256 points on the international large model arena "LMArena Vision Leaderboard," ranking fifth overall and first among domestic models, demonstrating outstanding performance. In multiple common academic evaluations on the OpenCompass multimodal academic evaluation set, it achieved an average score of 79.5, performing well in complex tasks such as visual reasoning, video understanding, and 3D spatial understanding.

📚 Features

Multimodal Input Support: Capable of processing various types of visual information including arbitrary-resolution images, videos, and 3D spatial data.
Strong Multilingual Capabilities: Significantly improved understanding of multilingual scenarios, accurately comprehending content described in different languages and performing analysis and processing effectively.
Technical Advantages
Advanced Architecture Design: The innovative MoE architecture balances computational efficiency and performance. Through a mixed routing strategy combining shared experts and specialized experts, it improves training efficiency and model performance.
High-Quality Data Support: A large amount of high-quality multimodal data was used during pre-training, including image and video data filtered and annotated through particular workflows, providing rich material for model learning.

📚 Application Scenarios

Photo-Based Problem Solving: Users can upload images, and the model will provide answers based on the content, such as identifying plants or solving math problems.
Video Calls: It can analyze video content in real time, providing relevant information or interaction, such as identifying the identities of call participants and analyzing scenes.
Video Understanding and Copywriting Creation: It can summarize and analyze videos, generate related copy, and create content based on video content, such as writing video scripts.