Open weights means an AI model whose trained files are published, so anyone can download it, run it on their own computer and change it. You get the finished model, but usually not the data or the code that made it.
Think of a cake. A closed model is a cake you can only eat at the bakery: they decide the price, the opening hours and the recipe changes. Open weights hand you the whole cake to take home, slice and decorate. What you usually do not get is the recipe and the shopping list, so you could not bake the same cake again from scratch.
In a sentence: “The model is open weights, so we downloaded it and run it on our own server instead of sending customer data to an outside company.”
What the “weights” are
The weights are the model. During training, an AI learns millions or billions of numbers, called weights or parameters. Together they decide what the model says. The US government’s telecoms agency describes the weights as “numerical values that the model has learned during training”.1
“Open” simply means the company put that file online for anyone to download, instead of keeping it on its own servers. NTIA also notes the catch: once a file is out, it is out.2
A simple example
Using a chatbot website is renting. Downloading open weights is owning a copy. With a closed model, your words travel to the company’s computers and the answer comes back. With an open-weights model, you download the file once, and a free program runs it on your own machine. Your chats stay on your computer, and the maker cannot change or withdraw your copy.
The trade-off: good models are large, so your computer needs plenty of memory, and answers are usually slower. Run AI on your own computer walks through it step by step.
Why it matters to you
Open weights give you independence, but the rules come from the licence. The file is yours to keep. What you may do with it is set by the licence published next to it. Some allow any use. Others forbid business use, or take rights away from large companies. One example: Mistral’s licence for one of its models cuts off companies above a revenue line.3
So before a business builds on an open model, someone should read that licence. AI labs outside the US and their open-weights licences compares the main ones.
Often confused with: open source
Open weights is not the same as open source. The Open Source Initiative, the group that keeps the definition of open source, says a truly open AI must also release the code used to train it and detailed information about its training data.4 Most downloadable models release the weights only, and many add limits on use.
| Word | What you get | What you can do |
|---|---|---|
| Closed | You get: access through a website or app | You can: use it on the company’s terms; it can change at any time |
| Open weights | You get: the trained model file | You can: run and adapt it, within its licence |
| Open source | You get: the file, the training code and data information | You can: also study and rebuild it, for any purpose |
AI terms people get wrong covers this mix-up in more detail.
Questions people ask
Is open weights the same as open source?
No. Under the Open Source Initiative’s definition, an open-source AI must also share its training code, the “complete source code used to train and run the system”, plus detailed information about its training data. Most downloadable models share only the weights.4
Is Llama open source?
Meta’s Llama models are downloadable, so they are open weights, but the Open Source Initiative says their licence does not meet its definition: “Llama 3.x is still not Open Source by any stretch of the imagination.”5 Always read the licence of the exact version you plan to use.
Can a company take back an open-weights model?
Not really. Copies that people have already downloaded stay out there. The US telecoms agency NTIA notes that “once weights have been widely released, it is difficult to ‘un-release’ them”.2
Learn more on this site
- Open or closed: who gets to hold the weights — the trade-offs for safety and power.
- Run AI on your own computer — how to try an open model yourself.
- Chinese AI models, explained — many well-known open models come from China.
- The AI glossary and AI for beginners — the basics, in plain words.
Related on this site: What Is a Checkpoint in AI? · Open Source vs Open Weights
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. What does open weights mean?
- The model’s trained files are published, so anyone can download, run and change it
- The model is free to use on the company’s website
- The company shares all its training data and code
Show the answer
Answer: The model’s trained files are published, so anyone can download, run and change it. You get the finished model to take home, like a whole cake, but usually not the recipe or the shopping list.
2. What decides what you may do with an open-weights model?
- Nothing, open weights means no rules
- The licence published next to the model
- The speed of your computer
Show the answer
Answer: The licence published next to the model. The file is yours to keep, but some licences forbid business use or take rights away from large companies.
3. How does open weights differ from open source?
- They mean exactly the same thing
- Open weights shares more than open source
- Open source also needs the training code and details of the training data
Show the answer
Answer: Open source also needs the training code and details of the training data. Most downloadable models release the weights only, which falls short of the Open Source Initiative’s definition.
Related terms: Latent space · Mixture of Experts (MoE) · Model collapse · Overfitting · Parameters (weights) · Pretraining
- 1NTIA (US Department of Commerce), Dual-Use Foundation Models with Widely Available Model Weights: Background, report published July 2024, read at source 1 Oct 2026“The weights represent numerical values that the model has learned during training to achieve an objective specified by the developers.” ↑
- 2NTIA, Background, report published July 2024, read at source 1 Oct 2026“once weights have been widely released, it is difficult to ‘un-release’ them”. ↑a“once weights have been widely released, it is difficult to ‘un-release’ them”. ↑b
- 3Mistral AI, Mistral Medium 3.5 licence, read at source 18 Sep 2026 (as cited on this site’s open-or-closed page): “You are not authorized to exercise any rights under this license if the global consolidated monthly revenue of your company (or that of your employer) exceeds $20 million”.↑
- 4Open Source Initiative, The Open Source AI Definition – 1.0, read at source 1 Oct 2026training code means “complete source code used to train and run the system”. ↑atraining code means “complete source code used to train and run the system”. ↑b
- 5Open Source Initiative, Meta’s LLaMa license is still not Open Source, 18 Feb 2025, read at source 12 Oct 2026“Llama 3.x is still not Open Source by any stretch of the imagination.” ↑
The comparisons on this page (the cake, renting versus owning a copy) are this site’s own ways of explaining, not claims from these sources.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK