HunyuanVideo 1.5 SSTA and FP8: or just call an API

HunyuanVideo-1.5 is an 8.3 billion parameter video model whose Selective Sliding Tile Attention (SSTA) gives a 1.87x speedup on 720p, and the project lists 14GB of VRAM as a minimum with model offloading enabled. If you do not want to operate GPUs, call a hosted model id instead and read its limits from the catalog.

What the repository states

The project page lists these facts.

What running it costs you

Open weights save the per-second price but move work to you: driver and CUDA versions, FP8 kernel support on your GPU, queueing, storage for outputs, retries and monitoring. A 1.87x speedup is relative to the baseline attention, not to a hosted service, so it does not tell you your cost per clip.

A fair comparison needs your own numbers: GPU hours per clip at your target resolution, divided by how many clips a day you make, plus the time you spend keeping the stack alive.

The hosted alternative on Sume

Sume exposes a Video Router catalog at GET /v1/video-router/models and the OpenRouter-compatible GET /v1/videos/models. Each model reports supported resolutions, aspect ratios, durations and pricing_skus, and limits differ per model: for instance wan-3.0 accepts 2 to 30 seconds and most other catalog models are capped at 15 seconds.

I did not find a HunyuanVideo id in the Sume docs I read, so check the live catalog if you need that specific model. Billing is list price times 1.25 on every Video Router model, and jobs are asynchronous.

A quick decision rule

• Run it yourself if you already own idle GPUs with at least 14GB (plus offloading) and need the weights.
• Call a hosted id if clips are occasional or you need more than one model family.
• Prototype on hosted first; self-host only if the volume justifies it.