Open-weight AI video models: what they are and when to use them

An open-weight model is one whose trained parameters, the weights, are published so anyone can download them and run the model on their own hardware. Earlier versions of Alibaba's Wan video model were released this way, which is why Wan comes up so often in conversations about running AI video yourself.

Open weights is not the same as open source in the full sense. The weights may be public while the training data and training code are not. And the license decides what you may do with the model and its output, which varies a lot from release to release.

A hosted model runs on someone else's servers. You send a prompt and get a video back. You never touch the weights. Most of the models in ClickPop work this way from your side: you pick a model, see the credit price, and render.

With open weights, you are the host. That brings real benefits and real work.

Video models are large and demanding. Running one locally usually needs a powerful graphics card with plenty of memory, or rented cloud GPUs. Check the model's official page for its hardware requirements before you start, since they differ by model and version.

Then there is setup. You download the weights, install the software it runs in, and often use a node-based or script-based workflow to connect the steps. Expect time spent on errors, version conflicts and slow first renders.

And there is maintenance. New versions arrive, tools update, and something that worked last month may need fixing. For a developer or a studio with a technical team, that can be worth it. For a creator who wants to post videos this week, it is often a detour.

Each open-weight release has its own license. Some allow broad commercial use, some restrict it, some require attribution, and some limit use by large companies. The terms can differ between versions of the same model family.

Before using open-weight output in ads, client work or monetized content, read the license on the official release page, and ask a lawyer if anything is unclear. Do not assume the rules of one version carry over to the next.

In ClickPop, videos you make are yours to use commercially, including ads and client work, as long as you avoid real people's likeness without permission and third-party brands or logos.

A two-person studio makes product videos for clients. Here is a checklist to decide whether to self-host:

If most answers point to hosted, use a studio with many models and one balance, and spend the saved time on creative work. If a specific need points to self-hosting, run that one model yourself and keep everything else hosted.

Self-hosting makes sense in a few clear situations. You need content to stay on your own machines for confidentiality. You want to fine-tune a model on your own licensed footage for a distinctive style. You are building a product where you need full control of the model version. Or you are a researcher studying how the models work.

It makes less sense when your goal is simply to make good videos quickly. Then the variety of models matters more than control over one. In ClickPop you can switch between 85 image and video models on one credit balance, render the same prompt on two of them and compare, which is hard to match with a single self-hosted model.

Many teams end up doing both: hosted tools for everyday production, a self-hosted model for one specific job.

If you want automation and control without managing GPUs, use an API. ClickPop has an API at /docs/api-reference and an MCP server at /docs/mcp, so scripts and agents like Claude can make videos across all the models with one credit balance.

That covers many reasons people look at open weights in the first place: batch production, custom pipelines and integration with other tools.