KIKOMONOWhy the word "open" misleads
An AI release is not one thing. It is a bundle of parts, and a publisher can release some parts while keeping others private. News coverage often shortens that bundle to a single word, "open," which hides exactly the parts a builder needs to know about. Two releases described with the same word can therefore come with very different rights and very different work.
The confusion is understandable: in traditional software, "open source" has a settled meaning — the source code is published under a licence that lets you use, study, change and share it. A trained AI model adds a part that traditional software does not have: the weights, the vast set of numbers produced by training. Publishing the weights is not the same as publishing the source code, and neither one automatically includes the training data.
The four parts, separated
| Part | What it is | What "published" does and does not mean |
|---|---|---|
| Weights | The trained numbers that make the model behave as it does | Downloadable weights let you run or adapt that trained model; they do not explain how it was trained |
| Code | The programs for training, running and evaluating the model | Published code lets you inspect and change the machinery; it does not by itself give you a trained model |
| Training data | The material the model was trained on | Rarely published in full; a description of the data is not the data itself |
| Licence | The legal terms for using, changing and sharing the parts above | The licence, not the headline, decides commercial use, redistribution and attribution |
A release can therefore be "open weights" while its code is unpublished, its data undisclosed and its licence restrictive — or it can be fully open on all four parts. The headline alone cannot tell you which one you are looking at.
Four checks before you trust the label
- Read the licence, not the announcement. Look for what it permits: personal use only, or commercial use too? May you redistribute the weights, or only use them yourself? Must you credit the publisher, and must modified versions carry the same terms? If the announcement does not link a licence, treat the release as unpriced, unlicensed territory: usable for reading about, not yet safe to build a product on.
- Ask what exactly is downloadable. Weights alone, weights plus code, or weights plus code plus evaluation tools? Each missing piece is work you must supply yourself, from writing the serving code to building your own tests.
- Separate "downloadable" from "free to run." Downloading weights costs nothing in itself, but running a large model needs hardware with enough memory, and that hardware is rented or bought. A model you can download is not a model you can run on an ordinary laptop, and a hosted version of the same model is a different product with its own price and terms. Compare the two before choosing, as in subscription versus API billing.
- Check whether "open" covers the version you actually want. Publishers sometimes release a smaller model openly while keeping the largest version behind an account or a paid plan. Confirm the size, capabilities and licence of the exact version you plan to use, not of the model family in general.
What this means for a small project
For a website builder, the practical question is rarely philosophical. It is: can I use this model in my project, under what terms, and at what running cost? A downloadable model gives you control — you can run it on your own server, keep prompts private and avoid per-request billing — but it also gives you responsibility: hosting, updates, safety checks and scaling are yours. A hosted model removes that work but keeps your project dependent on the provider's price, availability and terms.
Neither choice is wrong in general. The mistake is choosing on the strength of the word "open" and discovering the licence or the hardware bill afterwards. Read the licence first, confirm the exact version, and test with the workload you actually plan to run, following the model evaluation checklist.
For background on running models through hosted endpoints rather than your own server, see the inference endpoints guide.