Open-Source AI
AI Fundamentals
Open-source AI is AI released for anyone to use, study, modify, and share; for a model, that includes its weights, code, and training data details.
AI Fundamentals
What Is an Open-Weight Model?
An open-weight model is an AI model whose trained parameters, called weights, are published so people can download and run it themselves, subject to its license. You can host it on your own servers or a cloud you choose, instead of using it only through the developer’s service. Open-weight is not the same as open-source: many such models withhold their full training code or detailed information about their training data, and some licenses restrict certain uses.
Also called open-weights model, open-weight AI, open weights
Example
A mid-sized manufacturer stores years of engineering drawings, specifications, and test reports on its own file servers. Engineers want to ask plain-English questions such as “Which bracket designs passed vibration testing?” and get answers from those files. Some files contain confidential customer designs, so leadership does not want them sent to an outside AI service.
The IT team picks an open-weight model, confirms its license allows commercial use, and runs it on a powerful server in the company’s own data center. They connect it to a search index of text extracted from the specifications, test reports, and drawings, so the model draws its answers from those documents and lists the source files for engineers to check. Prompts and documents stay on the company network. In return, IT owns the hardware, security updates, and testing of each new model version.
How it works
An AI model is built from a huge set of numbers, called weights or parameters, that are adjusted during training. Together they hold what the model learned. When a developer publishes those weights, anyone with suitable hardware and software can run the model and get answers from it, a step known as inference, without going through the developer’s servers.
Weights are usually shared as downloadable files on model-sharing sites such as Hugging Face, sometimes only after you accept the license or request access. They typically come with a model card describing what the model is for and its known limits. From there you have several options:
- Run a smaller model on a well-equipped laptop or workstation.
- Run it on your own servers with one or more graphics processors (GPUs).
- Rent cloud hardware, or use a cloud service that hosts popular open-weight models for you.
Size drives cost. Small models can run on ordinary business hardware, while the largest typically need a dedicated server with multiple high-end data-center GPUs, or more than one such server. Many models are also available in compressed versions that need less memory, usually with some loss of quality.
Open-weight vs. open-source
A useful way to see the difference: the weights are the finished dish, while the training code and the details of the training data are the recipe. Many models described as “open” share the dish but not the recipe. You can use and adapt them, but you cannot fully inspect how they were made.
The stricter idea of open-source AI asks for the recipe too. The Open Source Initiative’s definition, for example, calls for the complete code used to train and run the model and enough detail about the training data for a skilled person to build a substantially equivalent system, under terms that let anyone use, study, change, and share it. Many popular open-weight models fall short of that, either because the recipe stays private or because the license restricts some uses.
In practice, vendors and news coverage use “open” loosely. Check what was actually released, and read the license, before assuming a model comes with open-source freedoms.
Why it matters for your business
With a proprietary model, you use the provider’s app or its API (a connection that lets other software send the model requests) on the provider’s terms. An open-weight model puts more of the decisions in your hands:
- Where it runs. On your own servers, in a private cloud account, or fully offline, which can matter for regulated or confidential work.
- Where your data goes. When you host the model yourself, prompts and documents do not need to leave your environment.
- How it behaves. You can adapt the model to your own examples through fine-tuning, which many proprietary models limit or do not offer.
- When it changes. If you host it, the weights stay the same until you choose to update them, so a vendor cannot quietly swap or retire your copy the way a hosted service can. Some licenses tie your use to policies the publisher can update, so recheck the terms from time to time.
- How you pay. If you host it yourself, costs follow your hardware, energy, and staff time rather than per-request fees. Cloud services that host open-weight models for you usually charge by usage instead.
None of this makes open-weight automatically the better choice. It moves control, and the work that comes with it, from the vendor to you.
Costs, risks, and what to check
Running a model yourself means taking on jobs a vendor would otherwise handle:
- Hardware and setup. Capable models need GPUs, plus people who can install, secure, and monitor them.
- Updates. You decide when to test and move to new model versions, and you patch the software around the model.
- Safe use. Hosted AI products often include content filters and usage policies. With a self-hosted model, adding guardrails, access controls, and logging is your responsibility.
- Quality. Performance varies widely between models, so test candidates on your own tasks before committing.
Read the license before building on any model. Check whether commercial use is allowed, whether certain uses are restricted, whether very large organizations face extra conditions, what attribution is required, and whether you may use the model’s outputs to train other models. Also note who published the model, since some organizations have policies about which sources they can use.
Open-weight vs. open-source vs. proprietary models
| Open-weight | Open-source | Proprietary | |
|---|---|---|---|
| What is released | Trained weights, usually with code to run them | Weights, training and running code, and training data details | No weights; you use it through an app or API |
| Can you run it yourself? | Yes, on your own hardware or a cloud you choose | Yes, on your own hardware or a cloud you choose | Generally no; it runs on the provider’s or its partners’ servers |
| Can you see how it was trained? | Usually only in part; training data and code are often withheld | Yes, in enough detail for a skilled team to study it or build a substantially equivalent model | No, beyond what the provider chooses to publish |
| Typical license terms | Varies, from permissive licenses to custom ones with use restrictions | Open licenses that allow use, study, changes, and sharing | Terms of service or a commercial contract |
| Customization | Fine-tuning and other changes to the weights | Fine-tuning, or retraining if you can gather the data and major computing power | Limited to options the provider offers |
| Who handles updates and safety | You, for the copy you run | You, for the copy you run | Mostly the provider, though you still own how you use it |
Every open-source model is also open-weight, but many models marketed as “open” are open-weight only, so the label alone does not tell you what was released or what the license allows.
Further reading
- AI ImplementationIf you are weighing whether to run a model on your own systems, this page describes SN3’s work redesigning processes, connecting systems, automating work, and deploying AI.
Beyond the definitions
Put AI to work for your business.
Go beyond the terminology. Explore practical AI training, workflows, and tools from SN3 Group.