Open-weight vs closed AI models, explained with a recipe
An open-weight AI model is published so you can download, run and adjust it yourself; a closed model stays with its maker. Explained with a published recipe vs a secret sauce.
- Author
Umair Mateen- Published
- Reading time
- 5 min
- Category
- AI
Two stories this week turn on one phrase: Microsoft says its Decision 1.1 model is built on Google's open-weight Gemma 4, while Alibaba says Qwen4-Max itself is not open weights.
In one sentence: An open-weight AI model is one whose trained settings (its "weights") are published so anyone can download it, run it on their own computers and adjust it, while a closed model stays on the maker's servers and you can only use it through their app or paid connection.
The everyday picture
Think of a famous restaurant dish. Some restaurants keep the sauce secret. You can order the dish as often as you like, at their price, but the recipe never leaves their kitchen. That is a closed model.
Other restaurants print the finished recipe card and hand it out. You can cook the dish in your own kitchen, change the spices, and serve it to your own guests. That is an open-weight model. The recipe card is the weights: the millions or billions of numbers the model learned during training, which decide its answers.
Notice what the card does not include: the chef's notebook of test batches and where every ingredient was bought. That is the training data and training code, and an open-weight release does not have to include them.
The lab publishes the trained weights. You download them, run the model on hardware you control, adjust it if you want, and the answers never leave your own kitchen.
How it works, step by step
- The AI lab trains the model (the chef perfects the dish). This takes huge amounts of data and computing power, which is why only a few companies do it.
- The lab publishes the weights (prints the recipe card) on download sites, together with a licence: the terms printed on the card saying what you may do with it. Google, for example, releases Gemma 4 under the Apache 2.0 licence.
- You run the model on hardware you control (you cook in your own kitchen): your own servers, a cloud account you rent, or for small models even a laptop or phone.
- You can adapt it (adjust the recipe). Teams fine-tune open weights for one job. Microsoft took Gemma 4 and trained it further to build its Decision model.
- Your data stays where you run it (the meal never leaves your kitchen). With a closed model, every question travels to the maker's servers instead.
Try it · Side by side
Published recipe or secret sauce: which suits your need?
Tap what matters to you (pick one or more) and see which option fits.
Open-weight model
- You first need hardware or a host to run it on.
- Can run on your own systems, so data need not leave.
- You can fine-tune the weights yourself.
- Your team or hosting partner handles them.
- Your copy keeps working, within its licence.
- Self-run hardware costs money even when idle.
Closed model
- Sign up, get a key and start asking questions.
- Each request goes to the provider's servers, under its terms.
- Limited to the options the provider offers.
- The provider runs and updates the model for you.
- The provider can change prices or retire versions.
- Pay per use, nothing to maintain.
Pick what matters to you and the better fit lights up.
Where the picture stops
- A recipe card is easy to read. Weights are just numbers. Even with them, nobody can simply read off why the model gives a certain answer.
- "Open weights" is not the same as "open source". The Open Source Initiative's definition of open source AI also asks for the training code and detailed information about the training data, so weights alone do not meet it.
- A published recipe can be copied but rarely misused. With AI, the US government agency NTIA notes that people with the weights can strip out safety features, and that a release cannot practically be taken back.
Why it matters for your business
- Privacy and control: running an open-weight model yourself means sensitive data need not be sent to an outside AI company. NTIA's report names this as a key benefit.
- Less lock-in: a model you downloaded keeps working the way it did, even if a vendor changes prices or retires a version. The licence terms still apply.
- More work for you: with open weights you run the kitchen. Hardware, updates, security and safety checks become your job, or your hosting partner's.
- Speed to start: closed models are usually the fastest way to try an idea. You sign up and pay per use, with nothing to install.
An example
Imagine a 30-person law firm that wants AI to sort incoming client documents. Its contracts say client files must stay on its own systems. A closed model would mean sending every file to an outside provider, so the firm tests a small open-weight model on a server in its own office. It accepts that its IT partner now has to keep that server patched and check the model's answers. For drafting marketing emails, where nothing is confidential, it keeps using a closed model through a normal subscription.
Frequently asked questions
Is an open-weight model the same as open source AI?
Not quite. Open weights means you get the trained model. The Open Source Initiative's definition also requires the training code and enough information about the training data for a skilled person to build a similar system.
Can I use open-weight models in a commercial product?
It depends on the licence attached to each model. Some, like Gemma 4 under Apache 2.0, allow commercial use; others add limits. Always read the licence before you build on one.
Do I need my own servers to use an open-weight model?
No. Many cloud providers host popular open-weight models and charge per use. Running them on your own hardware is an option, not a requirement.
Are open-weight models less safe?
They carry different risks. Anyone with the weights can remove built-in safeguards, but open weights also let outside researchers test and audit models. NTIA concluded in 2024 that the evidence did not yet justify restricting them, and recommended close monitoring.
Related explainers
The DevDrift helps teams choose between hosted and self-run AI models and put them into their workflows: see how.
Sources
Could you explain this to a colleague now?
Mark it and posts that mention it will show a tick. Saved in this browser only.