Decart, Amazon and the gap between open weights and free to use

On 11 October 2026 the Wall Street Journal reported that Amazon is in talks to acquire Decart, an AI startup whose models edit video, in a deal that could be valued at around $7 billion. The report cited people familiar with the matter and said a deal could happen soon if the talks do not fall apart or another suitor does not emerge. Reuters carried the report the same day. It said it could not immediately verify it, and that neither Decart nor Amazon responded to requests for comment.

Reuters also noted that Decart raised $300 million in May in a round led by Radical Ventures, with Nvidia joining as a new investor, and was valued at nearly $4 billion at the time, citing PitchBook data. Bloomberg had reported that Anthropic explored buying Decart in August for potentially $6 billion and, last month, decided not to pursue the deal after due diligence. For anyone who runs models on their own hardware, the price is the least interesting part. Decart is one of the few AI companies that published its video editing model as open weights, under a licence most people never read.

What it takes to run AI video editing locally with Lucy Edit Dev

Decart's best-known product is Lucy. According to the company's website, this Nvidia-backed model can edit live video in real time. That is the hosted product. Separately, Decart publishes a model you can download yourself: Lucy Edit Dev, which lives in Decart's organisation on Hugging Face.

Lucy Edit Dev does instruction-guided edits on video from text prompts. You give it a clip and a sentence, and it changes clothing or accessories, swaps a character, inserts an object or replaces the scene. The model card says it does this while preserving the motion and composition of the original footage. The person still walks the same way and the camera still moves the same way; only the thing you asked for changes.

Two more details matter if you want to try it. Decart also publishes a ComfyUI variant of the model, so you can load it into the same node-based workflow many people already use for image generation. And a later version, Lucy Edit 1.1 Dev, sits in the same Hugging Face organisation, so check which one a tutorial is using before you copy its settings.

This class of model is exactly the kind of workload people buy a homelab GPU for. Video editing models do the same sort of work as image diffusion models, but across many frames that have to stay consistent with each other, so they ask more of your hardware. I am not going to quote memory or speed figures here, because they depend on the version, the settings and the clip. The real requirements are on the model card and in the ComfyUI bundle. Read those first, then decide whether your card is a fit. If you are new to running models at home, my guide to open-weight models you can run yourself covers the basics of downloading, storing and loading them.

Open weights are not one thing: read the licence

Here is the part the headlines skip. The Lucy Edit Dev model card states its licence plainly as lucy-edit-dev-model-non-commercial-license. The weights are open, meaning you can download the actual file. The licence is non-commercial, meaning the file comes with rules about what you may do with it.

Decart's own catalogue shows how different two "open" models can be. Oasis is a world model that generates simulated environments, the kind used to train and test robotics and autonomous-driving systems. Decart built it with Etched, and its checkpoint, Etched/oasis-500m, is published under the MIT licence. MIT is a permissive open-source licence. It lets you use, modify and distribute the code and weights, including commercially, as long as you keep the copyright and licence notice.

So the same company has published one model you can build a business on and one you cannot, at least not without further permission. Both get called "open" in casual writing. Only one of them is open in the sense most developers mean.

In practice, this is what the difference looks like:

• A hobbyist editing their own holiday clips at home is usually on safe ground with a non-commercial licence. That is the use it was written for.
• A freelancer who uses Lucy Edit Dev to deliver paid work to a client is doing commercial work. A non-commercial licence is the wrong tool for that job unless the publisher grants a separate licence.
• A small business that wants to wrap the model in a product, or use it for its own marketing videos, faces the same question. Internal use for a company can still count as commercial, depending on how the licence defines it.

I am not a lawyer and this is not legal advice. The point is simpler: the licence text decides these questions, not the download button. Read the full licence file for how it defines commercial use, and if your income depends on the answer, ask the publisher in writing.

This nuance is almost invisible in news coverage. A headline that calls a model "open" tells you that a file exists. It tells you nothing about whether you can charge money for what you make with it.

What changes for you if the publisher gets acquired

Let me be clear about the status first. The Amazon deal is a reported negotiation, not a completed purchase, and neither company has commented. But the question is worth thinking through calmly, because it applies to every model you rely on, whoever publishes it.

The copy you already downloaded does not stop working. The weights on your disk are a file. They keep running on your hardware, under the licence they shipped with when you downloaded them. A change of ownership at the publisher does not reach into your server and rewrite that.

What can change is everything that comes next:

• Future versions might not be published as open weights at all. A new owner can decide the next release stays behind an API.
• Updated weights and fixes may stop appearing, so the version you have becomes the last one you get.
• Documentation, examples and community support can go quiet or move.
• A hosted API you pay for can change its terms, its price or its availability, because you never held a copy in the first place.

That is the real difference between owning a file and renting access. Owning a file under a restrictive licence still beats renting access you can lose, as long as your use fits the licence. Renting is convenient, but the terms belong to someone else.

A checklist before you build on any model

I run this check before I put any model into a workflow I care about. It takes a few minutes and saves a lot of rework later.

1. Read the licence file and what it permits commercially. Open the model card and the actual licence text, not a summary. Note the licence name exactly, such as lucy-edit-dev-model-non-commercial-license or MIT. Write down whether commercial use is allowed, and save a copy of the licence next to the weights.
2. Confirm the full weights are downloadable, and from more than one place. Some models publish only part of the system, or require an account and approval. Check that you can pull every file you need, and whether a mirror or the ComfyUI bundle gives you a second source.
3. Check that it fits your hardware. Use the requirements on the model card and in the ComfyUI bundle, not a forum post about a different version. Test on a short clip before you plan a project around it.
4. Look for a quantised or forked copy. If the community has made smaller quantised versions or maintained forks, a single publisher cannot switch you off. Forks inherit the original licence, so a non-commercial model stays non-commercial in every copy.

Once a model passes, store a local copy with its licence and a checksum, so you know exactly what you are running:

hf download <org>/<model> --local-dir ./models/<model>
cd ./models/<model>
sha256sum * > SHA256SUMS.txt

Back that folder up like any other data you would hate to lose. If the publisher's page changes tomorrow, you still have the file, the licence it came with and proof of which version it was.

If you want help setting this up

Running video models at home is one of the more rewarding things you can do with a GPU, and the licence check above is the step most people skip. If you would rather have someone set up ComfyUI, model storage and backups on your own hardware and harden the box properly, take a look at my services page.

Fiverr: hiteshsaini459 · Upwork: hiteshsaini25