Tencent released HunyuanVideo's weights under a community license

The shot

On 3 December 2024, Tencent released HunyuanVideo's weights, stating in its own GitHub repository that it was releasing "the inference code and model weights of HunyuanVideo," a text-to-video foundation model the documentation describes as having "over 13 billion parameters." Tencent's own framing calls this "the largest among all open-source models" at the time of release, a claim attributed here to the company rather than restated as independently confirmed.

What the documents show

The GitHub documentation and a corresponding Hugging Face model card both state the parameter count and describe the architecture: a transformer using full attention, a "Dual-stream to Single-stream" design, a multimodal large language model as text encoder, and a causal 3D VAE for compression. Tencent's own comparison states the model was evaluated against five closed-source models using "1,533 prompts evaluated by 60+ professional evaluators," with HunyuanVideo achieving the highest "Overall" ranking among the compared systems and, in the company's words, "excelling in motion quality." This is Tencent's own benchmark design and evaluator pool, not an independently run test, and no outside party's replication is cited. Both documents state the released weights require "an NVIDIA GPU with CUDA support," specifically a minimum of 60GB of GPU memory for 720p generation.

The workflow

A developer accesses the model by downloading the released weights and inference code from the repository and running it on qualifying hardware, rather than calling a hosted API, placing HunyuanVideo alongside other open-weight releases a studio's technical team could self-host or fine-tune. The documentation notes the repository was "tested only on Linux operating systems," a specific environment constraint for anyone planning a deployment. The Hugging Face card frames the model as usable for both image and video generation from a unified architecture, meaning the same weights serve more than one generation task.

What the tool does not change

Tencent's own documentation draws a distinction between the evaluated version and what ships by default: "the evaluation is based on Hunyuan Video's high-quality version," which it states "is different from the currently released fast version," meaning the benchmark results describe a variant a user may need to specifically select rather than the default behavior. The 60GB memory requirement means the tool does not remove the need for production-grade hardware planning; a studio still has to budget for infrastructure the documentation states plainly rather than treating the release as immediately usable on ordinary machines.

• Was the benchmark result being cited drawn from the "high-quality version" the documentation names, or the separate "fast version" it distinguishes?
• Does the available GPU infrastructure meet the stated 60GB minimum for 720p generation, or does the target resolution need to be reconsidered?
• Is Tencent's "largest open-source model" claim still current, or has a subsequent open release since exceeded it?

HunyuanVideo's release is documented as a scale and licensing event first: a large parameter count and a stated internal benchmark win, delivered as weights a team must provision serious hardware to run, rather than a plug-in tool with parity claims verified outside the company itself.