HunyuanVideo API: is there a hosted one, or only weights?

The Hugging Face cards read for this post show HunyuanVideo published by Tencent as downloadable weights with inference code; they describe no hosted API. Sume's docs do not list HunyuanVideo among the video models its API serves, so on Sume the answer is no: you either run the weights on your own GPU or use a model the catalog does list.

Tencent's facts below come from its Hugging Face model cards (HunyuanVideo-1.5 and HunyuanVideo), read 2026-09-29. The Sume side comes from Video generation.

What does Tencent publish?

The HunyuanVideo-1.5 card says the model has 8.3B parameters and that its authors released the inference code and model weights on Nov 20, 2025. Its checkpoint list has 480P and 720P text-to-video and image-to-video entries and distilled variants, with some still marked "Comming soon" on the card. The older HunyuanVideo card describes a video foundation model with weights and a single-GPU inference command.

Both cards carry the tencent-hunyuan-community license label. Read that license on Hugging Face before you build a product on the weights; this post does not summarize its terms.

What GPU does it need?

The two cards give different numbers, because they describe different models and different settings.

Is HunyuanVideo available on Sume?

No. The video catalog in code lists these ids: seedance-2.5, seedance-2-mini, seedance-2, seedance-2-fast, kling-3, wan-3.0, grok-imagine-video-1.5, minimax-h3, minimax-h3-max, gemini-omni-flash-1.1. HunyuanVideo is not among them, and the docs page on video generation does not mention it.

The catalog is the source of truth, so read it before you plan around a model. The public model catalog is how the docs suggest discovering video models:

curl "https://api.sume.com/v1/catalog" \
  -H "Authorization: Bearer $SUME_API_KEY"

Should I run the weights or use a hosted model?

Run the weights when you need the model itself: fine-tuning, an air-gapped machine, or full control of the pipeline. Use a hosted catalog model when you want a finished clip from one HTTPS request and no GPU to keep up. Open-source video model vs API walks through what each route changes, and Do you need a GPU for AI video? covers the hardware question.

What are the limits of this answer?

• It reflects what Sume's docs and code list on 2026-09-29. A catalog can change; check GET /v1/catalog.
• It says nothing about other companies' hosted HunyuanVideo services, because no vendor page for one was read.
• GPU numbers are the model authors' own and depend on resolution, frame count and offloading settings.