Open-weight model

Publishing the weights is not the same as open source. The training data and training code usually stay private, and the licence decides what you may do: some open-weight models ship under permissive licences such as MIT, while others carry bespoke terms with usage or revenue conditions.

Because anyone can run the weights, the same model is often available from many providers at very different prices and with different privacy terms, or on your own hardware. That makes the provider, not the model, the main decision when you choose how to run inference.

How it works

A trained model is, in the end, a large file of numbers. An open-weight release publishes that file, usually on Hugging Face, with a model card and a licence. Anyone with enough GPU memory can load it with an inference engine and serve it; providers do the same at scale and sell access per token.

The licence is the part to read. Permissive releases allow commercial use, modification and fine-tuning with few conditions. Bespoke licences can add revenue thresholds, use restrictions or rules about training other models on the outputs.

Example

A popular open-weight model is often available at the same time from its developer's own API, through marketplaces that route requests to many hosts, from private-inference services that run it on their own servers, and as weights to download and host yourself. The price per token for the same model can differ widely between those routes, and so can the privacy terms and context limits. That spread for one model is the practical meaning of open weights for a buyer.

In practice

If a developer suggests an open-weight model, ask three questions. Which licence does it carry, and does it allow your intended commercial use? Who will run it, a hosting provider or your own servers, and where are they located? And how was it compared with alternatives on your actual tasks? A model can score well in public tests and still write awkwardly in your language or format output poorly for your needs.

Providers serving the same open-weight model can differ in more than price. Some run reduced-precision versions to save cost, which can lower quality slightly; some offer shorter context limits than the model supports; and their data retention terms vary. Ask the provider for these details, or run the same test set through two of them. If you plan to self-host, include the full picture in the estimate: hardware or rented GPUs, monitoring, security updates and someone on call when it breaks. For most small businesses, a reputable hosted provider gives the benefits of open weights without that operational burden.

Why it matters for your business

Open weights give a business choices that closed models do not. You can pick a provider in a jurisdiction that suits your data-protection obligations under GDPR, switch providers without rewriting your product, or run the model yourself so customer data never leaves your servers. Competition between providers serving the same model also tends to push prices down. The trade-offs are real: self-hosting needs hardware and expertise most small companies lack, quality on some tasks can trail the best closed models, and you must check the licence. For many businesses the practical route is renting inference from a trusted provider and keeping the option to move.

Common mistakes

• Calling an open-weight model open source when its data, code or licence are not.
• Assuming every provider serving a model runs the same version, context limit and settings.
• Skipping the licence check before fine-tuning on, or commercially deploying, the weights.

Questions people ask

Are open-weight models free?

The weights are usually free to download, but running them is not. You pay either a provider per token or for the hardware, electricity and expertise to host the model yourself. Licences can also restrict commercial use above a certain number of users or for certain purposes, so free to download does not always mean free to use however you like.

Are open-weight models as good as closed models?

On many everyday tasks the gap is small, and on some the best open-weight models match closed ones. On the hardest reasoning or specialised work, leading closed models have often stayed ahead. The gap shifts with each release, so test the specific tasks you need rather than trusting general leaderboards.

Is it safe to use an open-weight model from another country?

The weights themselves do not send data anywhere; what matters is where the model runs. Hosted in your own environment or with a provider you trust, your data stays under those terms. Models can still reflect the biases or content rules of their training, so test answers on topics important to your business.