Llama 3.3 70B
Open weights
One OpenAI-compatible endpoint. Every model carries a machine-readable compliance record — where it runs, who can legally compel access to it, whether anything is retained, and which data classes it is fit for. Filter by your obligations, not by marketing claims.
Certifications shown are those held by the operator running the model.
Open weights
Open weights
Open weights
Open weights
Open weights
Open weights
Sending a prompt to a model is a data transfer. If that prompt contains personal data, client data or a trade secret, every rule that normally governs outsourcing and cross-border transfer applies to it — and most AI APIs are, legally speaking, a transfer to a foreign processor.
Switzerland's revised Federal Act on Data Protection, in force since September 2023. It requires a lawful basis for processing, a written agreement with any processor you use, and a record of processing activities. Transfers abroad are only allowed to countries the Federal Council recognises as adequate, or under safeguards such as standard contractual clauses. Sensitive personal data — health, religion, trade union membership, biometrics, criminal proceedings — carries stricter duties. Breaches must be reported to the FDPIC, and certain violations are criminal offences for the responsible individual, not just the company.
Applies whenever you process the personal data of people in the EU/EEA, regardless of where your company sits. It brings the same core duties as nFADP plus data subject rights, DPIAs for high-risk processing, and Chapter V rules on international transfer. Fines reach 4% of global annual turnover. Most Swiss companies with EU customers or staff are subject to both regimes at once.
A certification of the provider's information security management system, audited by an accredited third party. It tells you the operator has documented controls for access, encryption, incident response and supplier management, and that someone independent checked. It is evidence of diligence in your own supplier assessment — but note it says nothing about where data sits or who can compel access.
Residency is the physical location of processing. Jurisdiction is which government can compel disclosure — and the two are not the same. Under the US CLOUD Act, a US-incorporated provider can be ordered to produce data it controls even when that data sits in a Zurich or Frankfurt data centre. This is why a "Swiss region" from a US hyperscaler is not equivalent to a Swiss company operating its own infrastructure.
There is no single "compliant" answer. What you must insist on depends on your regulator, the data class, and how bad a compelled disclosure would be. A rough ladder:
Professional secrecy under Art. 321 SCC (lawyers, doctors, clergy), patient records, unreleased financial results, M&A material. Nothing leaves your premises, so there is no processor to contract with and no foreign authority to serve an order on.
FINMA-regulated banks and insurers under Circular 2018/3, health data, employee files, client identifying data. A Swiss company running its own data centres is subject only to Swiss law, so there is no CLOUD Act exposure to disclose in an outsourcing notification.
Listed companies and general corporate use where data stays in Switzerland but the operator is a US or other foreign group. Residency is satisfied; jurisdiction is not. Usually acceptable for internal and confidential data, rarely for regulated secrecy.
GDPR-adequate and often enough for EU personal data, marketing, and general business content. Note Switzerland is not in the EU: an EU host does not satisfy a Swiss-residency requirement, and a Swiss host does not by itself satisfy an EU-processing requirement.
Public and non-sensitive content only — published material, synthetic data, open-source code, throwaway drafts. Assume a foreign authority could compel access and that the prompt may be retained for abuse monitoring.
Zero retention means the prompt is not stored after the response. Some providers still keep 30 days for abuse monitoring, which is a disclosable sub-processing step. "No training on customer data" is a separate promise again — check both.
Open weights can be self-hosted, audited, and pinned to a version that will not change under you. That makes reproducibility and exit strategy realistic. Closed weights mean you depend on one vendor's continued availability and terms.
Local gives absolute control and no per-token cost, at the price of capability and uptime you must run yourself. Swiss cloud gives frontier-class quality with a contractual guarantee instead of a physical one.
The catalog filters map one-to-one onto the concepts above — residency level, sovereignty score, CLOUD Act exposure, training, DPA, data class. The guide explains how to set them for your company type.