What is an open-source model?
An open-source model is an AI model released with enough material for others to use, inspect, modify, and redistribute it under an open license. In the strict sense, that means the trained weights, the training and inference code, and a description or copy of the training data, all under terms such as Apache 2.0 or MIT that place no restrictions on who may use the model or for what. A few research groups publish models that meet this standard. Most models described as open source in vendor marketing are open-weight models, where the weights are downloadable but the training data stays private and the license carries conditions.
The distinction matters because the training data is where many legal and quality questions live. Without it, a company cannot verify what the model learned from, whether copyrighted or personal data was included, or why it behaves a certain way on a given input.
Why open-source models matter
For a business, an open model offers three things a closed model cannot. It can run on the company’s own hardware, so prompts and documents never leave the environment. It can be modified, through fine-tuning or LoRA adapters, to fit the company’s vocabulary and tasks. And it cannot be withdrawn, repriced, or silently changed by a vendor, because the files are already on the company’s servers. The cost is that the company, or a partner, must operate the model: provision GPUs, run an inference server, monitor quality, and apply updates.
Open-source models in practice
A county government wants an assistant that drafts responses to public records requests. Its counsel will not approve sending resident correspondence to an external API, and procurement rules favor software with no per-use fees. The IT team deploys an open model on a server in the county data center, fine-tunes it on past approved responses, and keeps every prompt and answer in an audit log the records officer can review.
What to watch for
- Read the actual license rather than the press release, and note any restrictions on commercial use, user counts, or fields of use.
- Check whether the training data is disclosed; a model card that describes it only in general terms signals an open-weight release.
- Confirm the model’s quality on the company’s own tasks, since open models vary widely.
- Budget for operations, including GPU capacity, updates, and the staff or partner who will run it.