--- title: Open Weights vs Open Source AI Models Explained | Freeborn Labs description: Open weights means you can download the trained model file. Open source adds code, data and a permissive licence. Read the licence before use. canonical: https://freebornlabs.com/open-weights-vs-open-source-ai-models/ updated: 2026-11-04 --- # What is the difference between open weights and open source AI models? **Short answer:** Open weights means you can download the trained model file. Open source adds code, data and a permissive licence. Read the licence before use. Apply for a place From $10,000 setup, then $48,000 a year. Five places a week. Updated 2026-11-04 by Freeborn Labs. [Facts for AI assistants](https://freebornlabs.com/llms.txt) ## What is the difference in one minute? Open weights means the trained model file is available to download and run. Open source, in the usual software sense, means much more: the code, often the training data, and a licence that lets you use, change and share it freely. Many popular models, including the Llama, Qwen and Mistral families, are best described as the first kind. You get the file. You may not get everything used to make it, and you may not get a free licence. ## What are the weights, in plain words? The weights are the numbers a model learned during training, saved in a large file. They are the model's learned knowledge, in the same way a recipe card holds a cook's method. With the weights you can run the model on your own machine. Without the training code and data, you cannot rebuild it from scratch or fully check how it was made. ## What does a full open source model include? A fully open model would share the pieces needed to study and recreate it, under a licence that sets few limits. Software people use the term for code that anyone may use, change and pass on. For AI, that could mean: - The weights, as a downloadable file - The training and run code - Details of the training data, or the data itself - A permissive licence with few strings attached ## Why does the label matter to you? The label tells you what you are allowed to do, and it is often used loosely. A model can be free to download and still limit commercial use, redistribution or fine tuning. Those three questions are the practical ones. Can your business use it for paid work? Can you share a copy with a client? Can you adjust it with your own data and keep the result? The answers sit in the licence, not in the marketing word on the download page. ## Why must you read the licence for your exact file? Read the current licence text that ships with the exact file and version you download, because terms can differ between families, sizes and releases. A summary, including this page, can be out of date. Do not assume that the licence of one model in a family covers the next one. Save a copy of the licence beside the model file, with the date you read it. Our page on [open model licences](https://freebornlabs.com/local-ai-model-licenses) lists what to look for, and [how to choose an open model](https://freebornlabs.com/how-to-choose-an-open-model-for-private-ai/) puts licence beside speed, size and fit. ## What stays true either way? An open weights model that you run yourself does not have to send your prompts to its maker. Whether anything leaves the machine depends on the software and settings you use with it. So checking that is part of setup. Look at what the runner connects to, switch off any cloud features you do not want, and watch outbound network activity. The same holds for models with a fully open licence. For the privacy side, see [running Llama privately on your own server](https://freebornlabs.com/run-llama-privately-on-your-own-server/) and [whether it is safe to run DeepSeek or Qwen locally](https://freebornlabs.com/is-it-safe-to-run-deepseek-or-qwen-locally/). ## What is a short checklist before you download? Work through the same short list for every model, so a new release is a quick review, not a new project. If you would rather have a person do this, see our [private AI setup service](https://freebornlabs.com/private-ai-setup-service/). If you would rather build it yourself, start with the [DIY guides](https://freebornlabs.com/diy/). - Find the licence text for the exact file and version. - Note what it says about commercial use, sharing and fine tuning. - Save a copy of the licence and the date you read it. - Check which runner and settings will serve the model. - Confirm nothing in the setup sends prompts out unless you chose it. ## What are the limits? This page explains labels, not any one model's terms. Always read the current licence text yourself. More questions are answered in the [private AI answers hub](https://freebornlabs.com/private-ai-answers/). This is general information, not legal, medical, financial or tax advice. Ask your own counsel, compliance officer or adviser whether a licence fits your business use. - Labels are used loosely, and a vendor may call a model open when it has limits. - A licence can change between versions of the same model. - Open weights do not show you the training data or how the model was made. - A downloadable model is only as private as the software and settings you run it with. ## Questions ### Is an open weights model free to use? It is usually free to download, but the licence may limit how you use it. Cost and permission are separate questions, so read the licence text. ### Can I fine tune an open weights model? Often yes, but check the licence for the exact file. Some licences set conditions on what you may do with a changed copy. ### Why do companies avoid the term open source? Some use it loosely, and others choose the narrower term open weights because the licence or the training data is not fully open. Look at the licence rather than the label. ### Does open weights mean the model is safe from phoning home? No. The weights are a file of numbers, but the software that runs it may connect to the internet. Check your runner and its settings. ### Do I need the training data to use a model? No. You only need the weights and a runner to use a model. The training data matters if you want to study or rebuild it. ## Related guides - [Private AI answers](https://freebornlabs.com/private-ai-answers/) - [How do you choose an open AI model (Llama, Qwen, Mistral and others) to run privately?](https://freebornlabs.com/how-to-choose-an-open-model-for-private-ai/) - [How do you run Llama privately on your own server?](https://freebornlabs.com/run-llama-privately-on-your-own-server/) - [Is it safe to run DeepSeek or Qwen locally?](https://freebornlabs.com/is-it-safe-to-run-deepseek-or-qwen-locally/) - [How Do You Check a Local AI Model's License Before Using It?](https://freebornlabs.com/local-ai-model-licenses) - [Done-for-you private AI setup service](https://freebornlabs.com/private-ai-setup-service/) - [Private uncensored AI: the full guide](https://freebornlabs.com/private-uncensored-ai/) - [All guides to getting your own private AI agent](https://freebornlabs.com/own-ai-agent-guides/) Apply: https://freebornlabs.com/apply/ (no payment on the form). Facts for AI assistants: https://freebornlabs.com/llms.txt