Tencent Hunyuan is Tencent's official hub for its Hunyuan (Hy) large models, covering language, vision, image, video, 3D and speech models, open-weight downloads, a free trial website for users in mainland China and paid API access through Tencent Cloud.
Used this tool? Rate it
Used this tool? Rate it
Tencent Hunyuan (腾讯混元; newer model names on the site use the shorter "Hy" prefix) is Tencent's official website for its Hunyuan family of large models. The privacy policy names Shenzhen Tencent Computer Systems Company Limited as the provider and describes a product built on Hunyuan and other large language models, covering text-to-text, text-to-image, text-to-video, image-to-text, image-to-image, image-to-video, video-to-video and audio-driven lip-sync video generation, plus tools and API access for creation, design, product development and content generation.
In practice the site is a catalog rather than a single app. The user agreement describes it as a place to try Hunyuan model features, read model news, and jump to other Hunyuan-based products and to API access. Each model has its own page, open-weight models link to their download repositories, and paid API use happens on Tencent Cloud.
This page covers the China site at hunyuan.tencent.com. Visitors from outside China are shown a prompt that suggests switching to a separate international site, which is not described here.
The catalog splits into models Tencent serves through its own products and API, and models it publishes as open weights. Several appear in both roles, so the same name can mean a hosted API, a downloadable checkpoint or both.
Most of these checkpoints link to download repositories from their model pages, but they do not share one license. Hunyuan open-source models range from Apache-2.0 releases to community licenses with territorial and user-count conditions, so the model choice and the license choice are made together (see Limitations).
The site's solutions section documents Hy inside Tencent's own products. These are vendor-published case descriptions, not independent audits, but they show where the models are deployed in production.
Product leads quoted on the site also give figures. WorkBuddy's context engineering lead reports that internal task success rose from 72% with Hy3 Preview to 90% with Hy3, with average completion time down 34%. A WeGame engineering lead reports that the Hy3-based Path of Exile 2 assistant reached 92% success on multi-turn reasoning and tool orchestration, with hallucinations falling from 4.5% to 2.8%. Both are internal measurements by Tencent teams.
Location changes the fit: the website and several model licenses carry geographic limits, which are listed under Limitations.
The website itself is currently free to use, with a cap on trial requests and possible queues when resources are tight; linked products and API access are charged according to their own pages.
API prices are published on the model pages and in Tencent Cloud's China documentation. Hy3 API pricing and Hy4 preview pricing are listed per million tokens:
The two Hy4 preview rows come from different language versions of the same model page and point to different Tencent Cloud consoles, so they are separate price lists rather than a currency conversion. On first activation, Hunyuan-a13b receives a one-time free package of 1 million tokens that expires one year after activation. Billing is by token usage, failed calls are not billed, and service does not switch to postpaid automatically when free or paid packages run out or expire.
Tencent names its own comparison models on the Hy4 preview and solutions pages. In a vendor-run blind test of 203 engineering tasks rated by 163 internal experts, Hy4 preview averaged 2.99 out of 4, against 2.92 for GLM 5.3 and 2.94 for Kimi K3. WorkBuddy reports that Hy3 used 47.4% fewer tokens than GLM-5.2 for document processing and 49% fewer for presentation creation.
Both comparisons are published by Tencent or its product teams. The directory and review-platform pages checked for this round offered no independent head-to-head comparison of these model families, so the practical differences to weigh are price per token, license terms and the hardware each option needs.
Three checkpoints use Apache-2.0: Hy3 (see Platforms), Hy4 preview and Hy-MT2. Tencent Hy4 preview is licensed under Apache-2.0. Hy-MT2 is licensed under the Apache License, Version 2.0.
Other checkpoints use a Hunyuan community license that does not apply in the European Union, United Kingdom and South Korea and adds a monthly-active-user threshold, measured on the model's release date, above which a separate license must be requested from Tencent:
The HunyuanImage 3.0 license also prohibits using the model or its outputs to improve any other AI model other than Tencent Hunyuan or its derivatives.
The website is currently free, with a cap on trial requests and possible queues. API calls on Tencent Cloud are paid by token after any free package, for example 1 RMB input and 4 RMB output per 1M tokens for Hy3.
The user agreement says the site currently serves users in mainland China only. Visitors from outside China are prompted to switch to a separate international site.
Among the licenses checked, Hy3, Hy4 preview and Hy-MT2 are Apache-2.0. HunyuanImage 3.0, HunyuanVideo 1.5, HunyuanOCR, Hy3 preview, Hunyuan 3D Omni, HY-World 2.0 and HY-Motion 1.0 use community licenses that exclude the EU, UK and South Korea.
The user agreement says uploaded and generated content is not used for model optimization unless you actively submit it as feedback. Inputs are still recorded to deliver the service.
It depends on the model: HunyuanVideo 1.5 lists 14 GB of GPU memory with offloading, HunyuanImage 3.0 recommends at least 3 × 80 GB, and Hy3 is recommended for 8 GPUs such as H20-3e.
