Open weights mean you get the finished numbers the model learned, not necessarily the training data or code. Licences vary widely: some allow any use, others restrict commercial use or large companies. Open-weight models can run locally or on your own servers, giving more control over privacy and cost.
How it works
Training a large model produces a file, or set of files, holding billions of learned numbers called weights. With a closed model, those files stay on the company's servers and you reach the model only through an app or API. With an open-weight model, the developer publishes the files, usually on a model hub, along with a licence and a model card describing training and intended use. You then load the weights into inference software on your own machine or a cloud server. Size decides the hardware. A model with a few billion parameters can run on a modern laptop, especially after quantisation, which stores weights at lower precision to save memory. Models with tens or hundreds of billions of parameters need one or more powerful GPUs. Because you hold the weights, you can fine-tune them on your own data, change system behaviour and run them fully offline. 'Open weight' is not the same as 'open source': a truly open-source model would also share training code and data details under a licence allowing broad reuse, which few large models do.
Why it matters
Open-weight models give you choice. A business can keep sensitive data on its own servers, avoid per-request API bills at high volume, and keep using a model even if a vendor changes or retires it. Developers can fine-tune a smaller open model to do one task cheaply and well. For creators and students, they make it possible to experiment, learn how models behave and build tools without paying for every call.
Where you will see it
- Running a private chatbot on company servers so documents never leave the network.
- Fine-tuning a small model to classify support tickets or write in a house style.
- Using AI offline on a laptop for drafting, coding help or study.
- Powering an app at high volume without per-token API charges.
Limits and common misconceptions
- Open weights do not mean free for any purpose; always read the licence before commercial use.
- You take on the hosting, updates, security and safety filtering that an API provider would normally handle.
- The largest open models need expensive GPUs, and small local models are usually less capable than leading closed models.
How it differs from related terms
- Open-weight model vs Foundation model A foundation model is any large general-purpose model; open-weight describes how it is released, and many foundation models are closed.
- Open-weight model vs Fine-tuning Fine-tuning is further training on your own data; open weights are what make fine-tuning possible outside the vendor's own platform.
Frequently asked questions
Is an open-weight model the same as open source?
Not quite. Open weight means the trained model files are released. Open source, in the traditional sense, also means the code, and ideally training data details, are shared under a licence that allows anyone to use, study, change and redistribute them. Many open-weight licences add restrictions that open-source definitions would not allow.
Can I run an open-weight model on my own computer?
Yes, if it is small enough. Models with a few billion parameters run on many modern laptops and desktops, especially in quantised form, using free local inference apps. Larger models need a strong GPU with plenty of memory, or a rented cloud server.
Are open-weight models less safe?
They carry different risks. Once released, weights cannot be recalled, and safety training can be removed by fine-tuning. On the other hand, open models can be inspected and tested by independent researchers. Whoever deploys an open model is responsible for adding suitable safeguards.
Related terms
- Neural network A machine learning model made of layers of simple connected units, artificial neurons, whose connection strengths are adjusted during training to learn patterns from data.
- No-code Building apps, websites or automations through visual editors, templates and settings instead of writing programming code.
- Parameters The numeric values inside an AI model, mainly weights and biases, that are adjusted during training and together store what it has learned.
- Prompt The text, files or other input you give an AI model to tell it what to do, including your question, context and instructions.
- Prompt engineering The practice of designing, testing and refining prompts and their surrounding context so an AI model gives accurate, consistent and useful output.