HunyuanVideo

What it is

HunyuanVideo is an open-source text-to-video generation model. Tencent said it had more than 13 billion parameters, which made it the largest open-source video generation model at the time ^[1]. Its weights are released under the Tencent Hunyuan Community license ^[3].

Origins

Tencent released the inference code and model weights on 3 December 2024 ^[1]. The technical report, titled HunyuanVideo: A Systematic Framework For Large Video Generative Models, was posted to arXiv the same day ^[2]. Diffusers integration followed on 17 December 2024, and FP8 weights on 18 December 2024 ^[1].

How it is built

The model uses a dual-stream-to-single-stream transformer with full attention. Video tokens and text tokens are first processed separately, and then concatenated for fusion ^[1]. Its text encoder is a decoder-only multimodal large language model, paired with a bidirectional token refiner.

Video is compressed by a causal 3D VAE, which reduces it 4 times in time, 8 times in space and 16 times in channels ^[1]. A VAE (variational autoencoder) is the component that maps video into a smaller latent representation and back, so the diffusion process works on the compact version rather than on every raw frame.

Hardware needs

The default output is a 720x1280 video of 129 frames ^[1]. Generating it needs at least 60GB of GPU memory, and an 80GB GPU is recommended ^[1]. Decrypt reported that running it locally takes at least 60GB of GPU memory, such as Nvidia H800 or H20 cards ^[4].

Evaluation

In a professional evaluation with 1,533 prompts and more than 60 evaluators, HunyuanVideo ranked first overall at 41.3%. It was compared with closed models including Gen-3 alpha (27.4%) and Luma 1.6 (24.8%). It led in motion quality with 66.5% ^[1]. These results come from Tencent's own evaluation.

Follow-up models

Follow-ups included HunyuanVideo-I2V, an image-to-video model released on 6 March 2025, HunyuanCustom on 9 May 2025, HunyuanVideo-Avatar on 28 May 2025 and HunyuanVideo-1.5 on 21 November 2025 ^[1].