Own your model: weights, data, evaluation and lineage that stay yours

You do not own the base model. An adapter or a fine-tune is a derivative of weights someone else released, under a licence they chose, and that licence still applies to what you build on it. Some licences are permissive and ask for little more than a notice. Others restrict use above a certain scale, require the model name to be shown, or exclude certain regions or uses. The Llama licence, for example, carries an EU restriction on some multimodal models, set out on the deploy-models hub.

So ownership starts before training, with the choice of base. Read the licence file that ships with the exact checkpoint, record the revision, and have counsel read it for your intended use. As a recommendation, permissive licences such as Apache 2.0 or MIT keep the most options open. If you used another model's outputs as training data, its terms may apply too.

A model you cannot take with you is a model you rent. Weights stored in the Model Vault can be downloaded, with the hash manifest that proves they are the files you evaluated, and run on any infrastructure that can serve them. Because Connect speaks the OpenAI-compatible API and so do the common serving engines, applications built against the gateway are not written against anything only Swfte provides.

Plan the exit while things are going well. Keep your datasets, evaluation sets and datasheets in storage you control, export the vault's audit log on a schedule, and test once that you can stand the model up elsewhere. The contractual side of exit and export is <contractual exit and export terms - founder to fill>.

Ownership is easier to defend when the infrastructure is also under your control or under law you understand. The platform's direction is sovereign by default: the company brain is customer-hosted and can run connected, private or air-gapped, and models are served on dedicated, single-tenant infrastructure rather than a shared pool. For European organisations, in-region hosting is the starting point; the regions and facilities on offer are <EU regions and facilities offered - founder to fill>.

Sovereignty has several dimensions, from where data sits to who operates the hardware to which law reaches the operator. The sovereignty page and the guide to deploying in the EU go through them one by one. For a custom model the practical test is simple: could you keep serving it if your relationship with any single provider ended tomorrow?

Lineage is the chain from data to model to decision. For a custom model it records the base and its revision, the dataset and its datasheet, the evaluation results at each promotion, and who approved. The Model Vault records the base model, versions, stages and an exportable audit log today; the dataset and evaluation records are yours to keep beside it until data export and managed evaluation are built.

Retirement matters as much as release. When a model is replaced, archive it rather than delete it, so the record of what served past decisions survives. Erasure is harder. If a person whose data was used for training later asks to be erased, current methods cannot reliably remove their influence from trained weights. The honest answer is to rebuild the dataset without them, retrain, evaluate and promote the new version, then retire the old one.

Owning a model costs more than calling one. You pay for the people who prepare data and build evaluation suites, for training compute, for dedicated serving that runs whether or not traffic arrives, and for retraining when data changes or is erased. In return you get behaviour you control, data that stays where you put it, and no dependency on a provider's next release. For a narrow, high-volume task that trade is often worth it. For occasional use, a governed general model is usually cheaper.

Swfte provides the technical controls, governance mechanisms and evidence required to deploy AI within an organisation's applicable regulatory, security and policy requirements. The exact posture depends on the customer's use case, jurisdiction, deployment and configuration. Swfte does not hold a SOC 2 report, an ISO 27001 certificate or a HIPAA BAA today. A SOC 2 Type I audit is in preparation; the trust page lists what is in place and what is in progress.

A public-sector agency adapts a small open-weight base, released under a permissive licence, to classify and route citizen enquiries in each of its official languages. It keeps the dataset, its datasheet and its evaluation sets in its own storage, and uploads the adapted weights to the Model Vault, which records the base model and hashes. Promotion to production needs sign-off from a named service owner, and the audit log is exported to the agency's records system each month.

When a citizen later asks to be erased, the team rebuilds the dataset, retrains and promotes the new version. When a procurement review asks whether the agency could move providers, it downloads the weights and stands them up on its own infrastructure as a test.