News

Releases & Models

Tencent launches Hy4 and says the open model enters the top tier of Chinese AI

Tencent launched Hy4, the new generation of its Hunyuan family, with first time computer vision and the same price as the previous version. The company says the model enters the top tier of open AI, but the weights have not been released yet.

Ilustracao editorial abstrata de uma esfera de rede neural vermelha e dourada abrindo um olho de camera, representando o modelo Hy4 da Tencent com visao computacional

Tencent officially joins the open model race

Tencent launched Hy4 on Wednesday, the new generation of its Hunyuan artificial intelligence model family, and the Chinese company was direct about it: it claims this is the first time one of its models breaks into the top tier of open weight systems, meaning models whose parameters anyone can download and run. The announcement came straight from Tencent's research blog, with immediate coverage from Nikkei Asia and the South China Morning Post, two outlets that closely track China's AI race. Hy4 arrives with a change that affects practical use right away: for the first time in the Hunyuan family, the model understands images, meaning it gained computer vision capability, able to analyze charts, documents and do visual reasoning, not just process text.

Aggressive pricing and immediate availability

The model is already live: Tencent Cloud, OpenRouter and Tencent's own products such as CodeBuddy, WorkBuddy, Yuanbao and ima all offer access to Hy4 from launch. The price charged for API usage is competitive even among Chinese rivals: 0.042 dollars per million cached input tokens, 0.834 dollars per million regular input tokens and 2.501 dollars per million output tokens, holding the same price as the previous version despite the jump in capability. This matters because the cost per token determines whether a company can run an AI assistant at scale without spending a fortune, and Tencent is competing head on with DeepSeek, Alibaba and Z.ai, which also sell cheap open Chinese models.

The numbers Tencent chose to show

Tencent reported that, in its own evaluation with 213 finance industry professionals, Hy4 came out ahead of open rivals on fund research, equity analysis and quantitative strategy tasks, scoring 3.04 against 2.97 for DeepSeek V4.1 Flash, 2.68 for Kimi K3.1 Pro and 2.68 for GLM 5.3. The company also says it cut the hallucination rate, when a model invents information that sounds real but is not, from 15.76% to 5.45% compared to the previous version. These are internal metrics, chosen and released by Tencent itself, so they should be read as the picture the company wants to show, not as an independent verdict. No outside lab has validated these numbers so far.

What is still missing to close this story

Tencent itself admits the limit: even breaking into what it calls the top tier of open models, Hy4 still trails leading proprietary models in advanced research level math, long horizon coding, self improvement capability and more complex visual questions. There is also a detail that matters for confirmation: the model weights, the file that lets anyone download and run Hy4 locally, have not been released yet. The company promises to open source them soon, has already partnered with the vLLM and SGLang communities for technical support, and says it will publish a detailed safety report alongside the weights. Until that release actually happens, public access to Hy4 depends only on Tencent's own paid API.

Sources

Tencent Hy: Introducing Hy4; Nikkei Asia: Tencent says its Hy4 joins top tier of open source AI models; South China Morning Post: Tencent unveils Hy4 as open weight AI race heats up.