Open-source AI vs open weights: what changes for you?

Why ‘open’, ‘free’ and ‘local’ are different claims

An announcement says a new model is open. A screenshot shows a free chat page. Someone else says they are running it on a laptop. Those may all describe the same release, but each answers a different question.

Free asks what you pay for a particular service. Open asks what is released and what you may do with it. Local asks where the computation happens. A hosted service can charge for running downloadable weights; a free website can keep its model private. Read the release and service terms separately.

This matters even if you never download a model. More options for running and adapting a system may lead to different products built around it. Your job as a reader is to understand which option became possible, not to assume that every ‘open’ announcement changes your own workflow.

What is the difference between open weights and open-source AI?

Weights are the learned numerical parameters of a model. Access to them is not, by itself, the same as access to the full process used to create the system.

The Open Source Initiative's Open Source AI Definition 1.0 requires freedoms to use, study, modify and share, along with the preferred form for making changes. That includes parameters, relevant code and sufficiently detailed training-data information. It does not simply equate openness with a downloadable file or require that every original training item be redistributed.

When a company uses the label differently, ask which materials and permissions it actually provides. We use the OSI definition as a named reference, not as a claim that all publishers use identical terminology. Source: Open Source Initiative: Open Source AI Definition 1.0

How do you check an open-model release without reading all the code?

Start with the model's own documentation. Hugging Face describes model cards as documents covering information such as intended uses, limitations, training and evaluation. A card is a place to begin, not proof that every statement in it is complete or correct.

Our suggested reading order is: identify the exact model and version; find the release files; read the intended use and limitations; then check the licence and access conditions linked by the publisher. ‘Available on a model hub’ is not a substitute for those checks. Source: Hugging Face: Model Cards

Use the full AI model-release checklist →

Can an open model be used commercially?

Check the actual terms attached to the exact release, rather than deciding from a headline. Hugging Face's licence documentation explains repository licensing metadata, which can help you locate the relevant information. A familiar model name or a download button is not the licence itself.

If you plan to redistribute or use a model commercially, have the applicable terms checked for your use. This article explains the labels; it does not determine your legal rights for a particular release. Source: Hugging Face: Licenses

Does an open AI model keep your data private?

Not automatically. The model's release terms do not tell you where an app sends your prompts. You might run downloaded weights on your own device or use a remote service running those same weights. A surrounding app may also connect to search, storage or other services.

For a concrete example, Ollama's FAQ distinguishes local use from cloud-hosted models and documents a local-only mode. That distinction is why you should inspect the actual setup rather than treating the tool's name as a privacy guarantee.

Ask where prompts are processed, what is logged and which external tools are enabled. Use public or fictional material while you are learning the setup. For workplace material, follow your organisation's approved tools and data rules. Source: Ollama: FAQ

Are open-weight models free to run?

A model may be available to download without a fee while still needing hardware, electricity, storage and time to operate. A provider running it for you may charge for that service. The useful comparison is the total effort and cost for your task, not just the download price.

For example, imagine you want help rewriting a short personal note. A hosted app may be convenient. If you want to experiment with a system you can configure yourself, a local setup may be more interesting. Neither example proves that local or hosted is always cheaper, better or appropriate for sensitive information.

Before changing your setup, write down the benefit you want: offline use, control over a configuration, an integration or a specific output improvement. If there is no concrete benefit yet, following the announcement can be enough. You do not owe every released model a weekend.

Is open-source AI better than closed AI?

Openness and task performance are different dimensions. A release can give you more access without being better at the thing you need. A polished hosted product can be useful without giving you the underlying weights.

Our recommendation is to compare two small examples of your actual task, with information you are allowed to share. Check whether the answer is right, what needed correcting and whether the setup met your requirements. Keep access, data handling and performance as separate lines in the comparison.

If an announcement leads with a benchmark, read its test conditions. If it leads with openness, read what was released. This keeps you from answering one question with evidence about a different one.

Which open AI developments are worth following?

Look for a meaningful change in what someone can do: a documented new capability, more useful release materials, clearer access or a practical way to run a task. Treat claims about freedom, affordability and quality separately until the evidence connects them.

Read The Day covers model releases and research in ordinary language, with links back to the source. If you want to understand why a release matters without collecting every model name, start with a sample briefing and keep the ones that are useful to you.

Keep up with AI without following every release →

Preview the free daily briefing →