Open Weights vs Open Source AI vs API-Only: Read the Label Before You Deploy

Three access models that should not share one label

These categories describe access and rights, not quality. A proprietary API may be capable and convenient. An open-weight model may be excellent for a controlled deployment. A system meeting an open-source definition may still require significant engineering. Procurement should compare the right dimensions instead of turning “open” into a universal score.

What model weights are

Weights are learned numerical parameters produced through training. Combined with a model architecture, tokenizer, configuration and inference code, they allow software to transform an input into an output. Making final weights available is meaningful: an organization may be able to archive a model version, run it without the publisher’s API and move between compatible serving environments.

Weights alone do not reveal the complete training process. They may not include dataset composition, filtering methods, training code, intermediate checkpoints or optimizer state. They also do not establish that every use is legally permitted. The repository’s licence and associated terms are part of the artifact.

What OSI’s Open Source AI Definition requires

The Open Source Initiative released version 1.0 of its definition in October 2024. It says an Open Source AI system must grant the freedoms to use the system for any purpose, study how it works, modify it and share it with or without modifications. Exercising those freedoms requires access to the preferred form for making modifications.

For machine-learning systems, OSI identifies three core categories in that preferred form: sufficiently detailed data information, the complete code used to train and run the system, and model parameters. The definition accounts for data that cannot legally be redistributed by requiring information about provenance, scope, selection and processing rather than pretending every underlying record can always be published.

This is a specific definition, not a casual synonym for “available on a model hub.” D-Central uses open weight when that is the claim supported by the release. We reserve open source AI for a system that can be evaluated against the stronger criteria.

A downloadable model can still have a custom licence

Some model releases use familiar permissive software licences. Others use model-specific community licences, acceptable-use conditions, field-of-use restrictions, redistribution obligations or thresholds that change rights for certain users. Terms may differ between code, weights, datasets and hosted services carrying a similar brand.

Do not infer permission from a download button, a repository tag or a blog post calling a model “open.” Review the licence linked to the exact repository revision, preserve a copy with the deployed artifact and obtain professional advice when rights are unclear. This page is technical procurement orientation, not legal advice.

Five questions that should remain separate

1. Can we obtain the weights? This determines whether self-hosting may be technically possible.
2. What are we allowed to do? Commercial use, modification, distribution and derived models depend on terms.
3. Can we study and reproduce the system? Code and data information affect transparency and meaningful modification.
4. Can our stack run it? Architecture, format, memory, runtime support and operational skill determine deployability.
5. Where will inference occur? Model origin and execution location are independent choices.

A model developed outside Canada can be served on Canadian-owned infrastructure if its licence and technical requirements allow it. A Canadian-developed model can be served by a foreign platform. Neither the passport of the publisher nor the location of a download page answers the sovereignty question.

Why open weights matter for a vendor exit

An API customer depends on the provider to maintain the endpoint, account, price, model version and supported jurisdiction. Downloadable weights can create another path: the organization or a Canadian operator can retain an approved version and serve it through a controlled interface. That can reduce concentration risk and make an exit plan testable.

Portability is not automatic. Applications may rely on a provider-specific schema, tool behaviour, embeddings, moderation service or context format. The alternative model may respond differently. A credible exit plan therefore includes a model-neutral interface where practical, exported data, versioned prompts, an evaluation set and a rehearsed cutover. The Canadian AI vendor exit plan addresses that operational work.

A licence review is only the beginning

Before deployment, record the model version, source, checksum, licence, notices, acceptable-use terms, supported runtime and intended use. Assess code and model-file supply-chain risk. Evaluate output quality and failure modes on representative tasks. Plan updates and decide whether new model versions require fresh approval.

Model benchmarks are dated evidence, not durable certification. D-Central’s open-weight AI comparison for Canada states its measurement date and separates licence, architecture and infrastructure fit from aggregate scores.

Frequently asked questions

Are all models on Hugging Face open source?

No. A hosting platform can distribute repositories under many different licences and access conditions. Read the exact model card, repository licence and linked terms.

Does open weight mean free for commercial use?

No. Access to files and permission for a particular commercial use are separate questions. Some licences are permissive; others impose conditions or restrictions.

Does open source guarantee a secure or accurate model?

No. Openness can improve inspection and autonomy, but security, provenance, evaluation, deployment controls and task fitness still require work.

Primary sources reviewed August 24, 2026: Open Source Initiative, Open Source AI Definition 1.0; Open Source Initiative, Open Weights; OSAID FAQ; Stanford HAI, Governing Open Foundation Models.