You download a weights file. It runs on your machine. The model answers your prompts and nothing leaves the building. It feels like freedom, and a lot of the time it is. But the word stamped on it, “open,” is doing more work than it can carry.
Open weights and open source sound like the same claim. They are not. One is about whether you can get the file. The other is about what you are allowed to do with it. The gap between those two is where people get surprised, and for a business it decides whether a model can go into something you sell.
What does “open weights” actually mean?
Open weights means the trained model is available to download and run. That is a real and useful thing. You get the actual file, you can run it offline, you can fine-tune it, and you are not renting access through an endpoint.
But “available to download” is not the same as “free to use however you like.” A weights file arrives with a license, and the license is the part that decides what you can actually do. Some open-weights licenses are genuinely permissive. Others are not. The download is identical either way, which is exactly why the word hides the difference. This is the same gap we flag when we check any AI tool before adopting it: read the license, not the headline. The word also appears before any file exists: Meta’s roadmap for Muse Spark promises an open weights release and gives no date for it.
That distinction is not pedantic. The whole reason to prefer a downloadable model is control: you can run it offline, keep your data on the machine, and adapt it to your own work. Every one of those benefits is real and survives even a restrictive license. What the license governs is the next step, turning that private use into something you ship, sell, or run at scale.
What does “open” quietly leave out?
Three things tend to go missing, and all three live in the license rather than the announcement.
Use restrictions. An open-weights model can ship with a license that limits what you are allowed to use it for. Certain uses are forbidden, certain industries are excluded, and past a certain scale you may need a separate agreement. That is a real constraint, and it is not what most people mean by open source.
The training data. Open weights gives you the finished model, not the recipe. The data it was trained on, the steps that produced it, and the code that did the training are frequently withheld. You can run the result. You cannot reproduce it or fully inspect how it came to be.
OSI-approved licensing. “Open source” has a specific meaning, defined by the Open Source Initiative, that includes the freedom to use the thing for any purpose. Plenty of open-weights licenses do not meet that bar. They are custom licenses written by the model’s makers, and they keep restrictions an OSI-approved license would not allow. A model can be downloadable and still not be open source in the established sense.
None of these three is hidden maliciously. They are simply not what the word advertises. “Open” now spans everything from a fully OSI-approved license to a downloadable file wrapped in pages of restrictions, and the only way to know which one you are holding is to read past the headline before you build on it.
When does the model license actually matter?
For a lot of work, none of this bites. You run the model, it does the job, and the license never enters the room.
It enters the room the moment money or scale shows up. If you are building a product on top of a model, putting it in front of clients, or running it at volume, the license is not a formality. It decides whether you are allowed to do the thing you just built. A clause you skimmed can be the difference between a tool you own and a tool you have to tear out later. This is why license terms sit alongside speed and quality when we decide which local model to actually run, not as an afterthought.
A concrete example helps. Two models post identical benchmark scores. One ships under a permissive license that lets you build commercially without asking. The other is downloadable but caps commercial use above a revenue threshold and bans a few categories outright. On raw quality they tie. For a business billing clients, they are not remotely the same model, and the difference is invisible until you read the file.
The permissive case does happen. Google’s DiffusionGemma release ships under Apache 2.0, a licence that allows commercial use, and its headline claim is text generation up to four times faster than conventional models.
The asymmetry is the trap. The download button looks the same, the model card reads the same, and the benchmark table treats the two as equals. Only the license separates them, and it is the one document nobody screenshots for the launch thread.
Open weights vs open source for a business: can you ship on the models we run?
The table below covers four models we have run on our test machine, an M4 Pro with 48 GB of unified memory, and what each licence says about commercial use. Each licence was read on 17 October 2026 from the maker’s own page. This is a reading of the documents, not legal advice, and a business shipping a product should have its own adviser read the same pages.
| Model | Where we ran it | Licence | Can a business ship on it? | User or revenue cap | What you must keep |
|---|---|---|---|---|---|
| Gemma 4 31B | Our local text model | Apache 2.0, per Google’s Gemma 4 licence page | Yes | None in the licence | A copy of the licence, notices, and a mark on changed files. Google’s site also links a separate prohibited use policy |
| Qwen3.6 35B-A3B | Our parallel request test | Apache 2.0 | Yes | None in the licence | A copy of the licence, notices, and a mark on changed files |
| FLUX.1 schnell | Our ComfyUI image pipeline | Apache 2.0 | Yes; the model card says personal, scientific and commercial purposes | None in the licence | A copy of the licence and notices; the model card lists out-of-scope uses |
| LTX-Video 2B 0.9.8 distilled | Our LTX-Video test | LTX-Video Open Weights License | Yes, with conditions | Entities with annual revenue of at least $10,000,000 need a separate paid commercial licence | A copy of the licence, marked changes, notices, the use restrictions passed on, and machine-generated content disclosed as such |
Three of the four use Apache 2.0, an OSI-approved licence, so their licence terms meet the open source bar described above, whatever their makers withhold about training data. The fourth is the kind of licence this piece warns about. LTX-Video is free to download and free to run, but a company over the revenue line needs a separate agreement, and its attached list of banned uses includes medical advice and content that is not disclosed as machine-generated. A business choosing between a video model and an image model on output quality alone would miss that.
Is open weights still worth it?
Open weights is still a real gift, and this is not a reason to dismiss it. A downloadable model you can run offline and fine-tune is worlds better than an API you can only rent. The freedom to inspect, to run privately, and to not depend on a vendor’s uptime is genuine, even under a restrictive license. It is the same freedom that lets you prove your data never leaves the machine. Plenty of teams will never hit the limits in the fine print, and for them the distinction is academic.
The point is smaller and sharper than “open weights is bad.” It is: do not let the word “open” do your reading for you. The license is a short document, and it is the only place the real terms are written down.
Next time you pull a model, open its license before its first prompt. Read the part about what you may and may not do. The five minutes it takes is the cheapest insurance you will buy on the whole project, and the difference between a tool you own outright and one that is borrowed on terms you never checked. On a small project that gap is a footnote. On the one that finally takes off, it is the whole story.
