Vibeleaderboard
Tencent

ai lab

Tencent

Tencent matters because Hunyuan and Hy sit inside a company with cloud infrastructure, games, communication, productivity, advertising, and content products. It can test and distribute models across many real workflows while also releasing useful video and 3D artifacts to developers. The unresolved question is whether current general-model progress will gain the independent evaluation and international developer adoption already visible for China's strongest open model labs.3,5,7,11,12

Official site ↗Models in the Atlas →

Profile

Overview

A research program inside a product conglomerate

Tencent is the Shenzhen technology company founded in 1998 and best known for WeChat, QQ, games, cloud services, and digital content. Its corporate AI research predates the current foundation-model cycle: Tencent AI Lab was established in April 2016 to work across machine learning, computer vision, speech, and natural-language processing. The company later organized a broader laboratory matrix that included AI Lab, VisionSeed, and Weixin AI.2,1

Hunyuan enters Tencent Cloud

Tencent launched its Hunyuan foundation model for enterprise access through Tencent Cloud in September 2023. At launch, the company said Hunyuan was already connected to more than 50 internal products, including Tencent Meeting, Tencent Docs, Weixin Search, games, marketing, and cloud services. The program mixes proprietary services with downloadable releases, so Tencent's model reach cannot be measured only through public weights or a single assistant.3,4

Open work in video and 3D

The most visible open research is multimodal. HunyuanVideo released code and weights for a large video diffusion transformer, then expanded into image-to-video, avatar, custom-video, and lighter 1.5 variants. Hunyuan3D 2.0 separated shape generation from texture synthesis and released models for high-resolution textured assets. These reports document substantial open artifacts, but their headline comparisons largely use Tencent-run evaluations and the community licenses impose regional and use conditions that are not equivalent to permissive open-source software licenses.5,6,7,8

The Hy3 rebuild

Tencent rebuilt its general-model pretraining and reinforcement-learning infrastructure in early 2026 under chief AI scientist Yao Shunyu, previewed Hy3 in April, and released it in July. Tencent says the model now supports Yuanbao, WorkBuddy, Marvis, games, customer service, and Weixin workflows. Reuters reported that second-quarter revenue rose 11 percent while AI infrastructure spending increased. That is evidence of deployment and investment, not independent proof that Hy3 matches the strongest Chinese or global general models.9,10,11

Notable contributions

  1. 01A broad open Hunyuan media stackTencent released code and weights across video, image, 3D, avatar, and related generation workflows. The contribution is the breadth and usability of the model stack, not a claim that Tencent invented diffusion-based media generation.5,6,7
  2. 02Two-stage high-resolution 3D asset generationHunyuan3D 2.0 paired a diffusion-transformer shape model with a separate texture-synthesis model and released inference code and weights, supporting downstream 3D asset workflows while preserving a clear geometry-texture boundary.7,8
  3. 03Foundation models embedded across a product matrixTencent connected Hunyuan to cloud APIs and dozens of company products, then used product feedback in the Hy3 rebuild. This is a deployment-system contribution rather than a claim of architectural priority.3,9,10