Open-weights video or hosted API? Four checks, Wan, LTX, Runway

Choose open weights when you need control over the model, data path or fine-tuning and can run GPUs; choose a hosted API when you want clips on demand without operating infrastructure. Four checks decide it: hardware, license, price per clip, and who runs the service. The vendor pages below give the real numbers for Wan2.2, LTX-2.5, Luma and Runway.

Check one: hardware

The Wan2.2 README (read 2026-10-02) says the TI2V-5B model needs at least 24 GB of VRAM and the 14B models want 80 GB or more on one GPU, or multiple GPUs with FSDP. The LTX-2.5 card (read 2026-10-02) lists a 22B transformer and recommends Python 3.12, CUDA 12.7 and PyTorch about 2.7.

Check two: license

Wan2.2 is listed under Apache 2.0 on its repository. LTX-2.5 is free for commercial use under $10M annual revenue under the LTX-2.x Community License, with a paid agreement above that. Open weights are not all open on the same terms, so read each one.

Check three: price per clip

For self-hosting, price is GPU hours divided by clips per hour. Measure clips per hour on your own prompts before you trust any estimate.

Check four: who runs it

Hosted APIs bring queues, storage, retries and moderation. Sume jobs are asynchronous: submit to POST /v1/videos, poll the job, then download the result, per the video docs. Self-hosting means building those yourself.

Sume's catalog does not include Wan2.2 or LTX weights; it hosts catalog models such as Wan 3.0 (2 to 30 seconds per clip).

A rule of thumb

Recheck every number here on the vendor page before you decide.

• Prototype on a hosted API; the first clips cost cents, not a GPU contract.
• Move to open weights when volume, privacy or customization justify the operations work.
• Keep your prompts and references portable so you can switch.