HunyuanVideo 1.5 on 14 GB of VRAM: run locally or call an API

The HunyuanVideo 1.5 repository, read 2026-10-03, lists 8.3 billion parameters, output at 480p, 720p and 1080p, and a minimum of 14 GB of VRAM when offloading is on. A step-distilled variant runs in 8 to 12 steps. The repo material dates from December 2025, so it is older than the other items in this series.

The specs

These come from the repository page.

What 14 GB means in practice

14 GB is a floor with offloading enabled, which trades memory for time by moving parts of the model in and out of GPU memory. The repo figure does not tell you how long a clip takes, so measure that on your own card before you plan around it.

A local versus API checklist

Answer these before you pick.

• Do you need to keep inputs on your own hardware? That favors local.
• Do you need many clips in parallel? An API job queue handles concurrency; on Sume, valid paid jobs are accepted as queued while queue capacity remains.
• Do you need audio? Check the model card. Sume reports an audio flag per catalog model.
• Who handles failures and retries? Locally, you. On Sume, you send an Idempotency-Key and poll the stored job id.
• What is the cost per finished clip? Local cost is your hardware and time; hosted cost is the per-second rate. Sume bills provider list times 1.25.

Measure before you plan

Run a five-clip test on your own hardware and record three numbers: seconds to finish, peak VRAM, and the number of clips you would keep. Divide your GPU cost per hour by clips kept per hour to get a cost per finished clip, then compare it with a hosted per-second rate times your clip length.

Date the test. Drivers, offloading settings and repo updates change the result.

A hybrid path

A common pattern is to use the local model for drafts and a hosted model for finals, or the reverse. Keep one small wrapper in your code that takes a prompt and returns a file, and swap the implementation behind it. With Sume, the wrapper is a submit to POST /v1/videos, a poll on the returned polling_url, and a download from unsigned_urls[0], as described in the video generation guide.