Skip to content

About

Sovereign AI shouldn't mean second-best AI.

sysf.io exists to make that true in practice: an open Swiss model, fine-tuned on your documents and measured on your task before it ships — not a locked-down black box in someone else's datacentre.

Why we exist

The sovereignty gap.

The most capable AI has, until recently, lived behind closed US APIs. For a Swiss bank, a hospital or a cantonal authority, that's a dead end: the data can't leave, so the pilot dies.

Switzerland answered with Apertus — a fully open, sovereign, multilingual model from EPFL, ETH Zurich and CSCS, with open weights and documented training data. It closes the capability gap while keeping everything inspectable and ownable — the measured proof (Ticino's 90→94% translation service, our FINMA ablation) lives under News.

sysf.io is an independent Swiss engineering studio that turns that open foundation into working systems. We do the hard part — fine-tuning, retrieval, evaluation, compliance documentation and operations — so you get a system you own, measured on your task — with the evaluation report to show your risk committee.

The name

systems, fine-tuned.

The name says what we do. No cloud, no black box — just your systems, fine-tuned.

sysf the name, decoded
sys
your systems — data, documents and workflows, inside your walls.
f
fine-tuned — an open Swiss model tuned to them, measured before it ships.

Principles

How we work.

Sovereignty first

Training and inference run in your tenancy or your datacentre; Production and Air-gapped engagements have zero data egress.

Open over closed

Open weights only — Apertus, Llama or Mistral, whichever wins your eval — so auditors can inspect what they sign off.

Products, not demos

Go-live is gated on an evaluation score agreed at scoping, not on a stakeholder demo.

Compliance by design

FADP/GDPR mapping and EU AI Act documentation are named deliverables in the statement of work.

How this compares to hosted platforms and US cloud APIs — see the comparison on the platform page.

Work with us.

Bring one workflow and the data constraint that blocks it. You'll get a 30-minute call with the engineer who'd build it — and a clear no by minute 20 if we're the wrong fit.