AI video for brands: local models or cloud APIs in 2026

Wan 3.0 provides a useful reference point for comparing two approaches to AI video production: running open models locally or generating clips through a cloud API. Alibaba introduced the public beta of Wan 3.0 on 7 August 2026 and announced general availability on 24 August 2026. The model supports clips of up to 30 seconds, but it is offered through an API and its weights have not been released. The available open-weight line remains Wan 2.2.

For a brand, this means you should not confuse access to a new model with the ability to run it on your own equipment. Your decision depends on file confidentiality, production costs and the level of control you need over each result.

What are your options for AI video production?

You can organise the decision around three approaches. Each addresses different needs and requires you to consider both visible costs and technical production requirements.

A local model may suit you if you need to keep reference materials within your own infrastructure. An API can reduce the need to set up your own generation environment and provide access to a newer version. A hybrid approach lets you separate sensitive assets from tasks that can be outsourced.

How much does Wan 3.0 video generation cost?

Published Wan 3.0 pricing depends on resolution. According to the available information, the service supports generation at 480p, 720p and 1080p at 30 fps, with the following per-second costs:

These figures help you calculate the base cost of a generation, but they should not be confused with the full budget for an advertising asset. If you need several versions, scene corrections or repeated generations, total spend will depend on the number of attempts and the duration of the clips you produce.

You should also avoid converting these amounts directly into euros for a project budget without checking the exchange rate and applicable billing terms.

What do you gain and lose by working locally?

The main advantage of running a model on your own equipment is control over where your assets are processed. If you work with product prototypes, confidential designs or campaigns that have not yet launched, this can be a decisive consideration.

The trade-off is that you are limited to the models you can run and the resources available on your equipment. In this case, open weights are available up to Wan 2.2, under the Apache 2.0 licence according to the available information. Wan 3.0, by contrast, has no published weights and is accessed through an API.

Before choosing a local option, check the licence of each component you plan to use and the terms governing commercial use. Open model weights do not remove the need to review the relevant terms or guarantee that every combination of tools is suitable for your production.

You can extend this review with our analysis of what remains after Sora closes and which licences apply.

What does a cloud API offer, and when does it make sense?

Wan 3.0 accepts text, image, video, audio and document inputs, including PDF and PowerPoint files. This range allows you to consider workflows that start from different types of source material, although the specific options depend on the feature you use.

An API can be useful when you need access to a particular model version without maintaining its weights or preparing the infrastructure to run it. In return, you need to account for generation costs and check what happens to the files you submit.

Before using client assets, review confidentiality, retention and data-handling terms. Do not assume that a cloud service offers the same safeguards as a local environment. Equally, do not assume it is available in a particular region without checking the provider's terms.

How can you build a hybrid workflow without losing control?

A hybrid workflow can reserve local processing for sensitive material while using the cloud for tasks that do not require confidential information to be uploaded. To decide whether this approach makes sense, map the production stages and classify your files before selecting tools.

Start with these steps:

• Separate public assets from files subject to confidentiality requirements.
• Decide which tasks require local generation and which can run through an API.
• Calculate the cost per clip, including the repetitions you expect to need.
• Check the licences and commercial-use terms for each component.
• Validate the visual result before producing a full sequence.

If you need to deliver a sequence for compositing or post-production, define the output format from the start as well. Our guide to generative video delivered in EXR explains why the output needs to suit the workflow that will handle the footage.

What decision should you make this week?

This week, choose a representative video brief and compare a local test with an API generation. Use the same visual objective, record the cost of each attempt and check which assets you can share with each tool. This will help you decide which parts of the process should remain within your infrastructure and where paying for cloud generation makes sense, without assuming that one solution fits every project.

If you want to produce AI-generated product videos or campaigns and establish a workflow suited to your assets, tell us what you need to create and which confidentiality requirements you need to meet.