
Video generation is the most expensive category to iterate in: a prompt costs real money, a retry doubles it, and a model update can change the look of everything you produced last month. Open weights change that arithmetic, and they add a different one.
This page is about Wan as the open-weight option, against hosted tools like Sora and Hailuo.
The short answer
Open weights buy a pinned model, on-premises inference, and no per-clip fee. They cost GPUs, a serving stack, and the evaluation work that a hosted product does for you. Choose them when the model version is a compliance or consistency requirement, not to save money at low volume.
What a pinned model changes
| Property | Hosted | Open weights |
|---|---|---|
| Version stability | Changes under you | Pinned |
| Cost shape | Per generation | GPU hours |
| Data location | Vendor | Your hardware |
| Quality ceiling | Frontier | One or two steps behind |
| Ops burden | None | Yours |
The first and fourth rows are the real trade. A consistent look across months is worth a lot for a brand; a frontier quality lead matters for a one-off campaign.
Wan in practice
Wan is a family of open-weight video models with image-to-video and text-to-video variants. The practical use is a fixed pipeline: one model version, one set of parameters, and a queue that a product can call. It is not a better model than the hosted frontier, and it is a model you can point at.
The evaluation problem
Hosted video models are evaluated by trying them. An open-weight model has to be evaluated by you: prompt adherence, motion coherence at three and eight seconds, and the failure modes that matter for your shots. That is a day of work per model version, and it is the hidden cost.
The cost crossover
At a few clips a week, hosted is cheaper once engineering is counted. At a thousand clips a month, or for a product that generates video per user, the GPU math starts to win. The crossover is the number to calculate before committing, not after.
Where the pipeline sits
The model is one step. Shot lists, retries, captions, and crops are the rest, and they do not change with the model. Teams that invest in the pipeline get more out of a slightly weaker model than teams that chase the frontier and re-do their workflow every update.
Frequently asked questions
Can I fine-tune a video model? Some open-weight families support it; the data requirement is much larger than for image models.
What about licensing? Read the weights' licence; it is separate from a hosted product's terms. See licensing AI images.
Is Wan better than Sora? No, and it is not trying to be. See Wan vs Sora.


