Open weights

Open weights AI is a model whose trained parameters are publicly available, but not necessarily under an OSI-approved open source license. Distinct from open source AI.

Open weights vs open source AI

An open weights model publishes the trained parameter file (the weights) and typically inference code, but the training data, training code, and license may not meet open-source criteria. Meta's Llama series is the canonical example — weights are downloadable, but the license restricts commercial use above a user-count threshold and the training data is not released.

An open source AI model, per the Open Source Initiative's OSAID 1.0 definition (October 2024), requires sufficient information for someone else to recreate a substantially equivalent system, including data information, training code, and weights under OSI-approved licenses. Few production AI systems currently meet this bar.

Why the distinction matters

Marketing and casual use treat "open" as a single concept; license teams treat it as several distinct legal regimes. Open weights enables self-hosting and customization but does not eliminate license obligations. Open source AI (in the strict OSI sense) gives the broadest freedoms but is currently rare for frontier models. For procurement, the open-weights vs open-source distinction determines whether you can fork, modify, and redistribute the model under your own terms — significant for jurisdiction-sensitive deployments.

Buyer-side checks

For self-hosted "open" AI deployments: which license applies (Llama Community License, Apache 2.0, MIT, custom), what are the commercial-use thresholds, are there acceptable-use clauses that bind your downstream applications, what data licensing covers the training corpus, and is there a path to OSI-approved open source if you need it. The "open" label without specifics is rarely sufficient for enterprise procurement.