Made for Swiss self-employed & SMEs
Every week brings a new AI service with a price tag, and for Swiss SMEs and self-employed professionals, the costs can add up fast. Yet many routine tasks — routing client emails, categorising invoices, prioritising support tickets — only need simple classification, not a full-blown model. Kev is an open-source classification API you can run yourself, with sizes from 0.8B to 27B parameters, where the 4B model is the sweet spot that runs on a laptop GPU or even CPU for basic workloads.
\n\nWhat Kev actually does
\n\nKev answers yes/no, multiple-choice, or rating questions about a piece of text and returns calibrated probabilities in one pass. It is built for ticket routing, moderation, and classification — not free-text generation. Think of it as a wire-compatible drop-in replacement for paid services like Jev (the closed model behind TypeSafe's System One API), except you host it yourself or run the 4B model locally on a machine you own.
\n\nThree tasks that usually pay off first
\n\nIn our work with Swiss freelancers, notaries, doctors and small offices, three classification tasks almost always repay the effort:
\n\n- \n
- Ticket routing. Incoming client emails or support requests can be automatically directed to the right team or flagged for follow-up based on content keywords. \n
- Document categorisation. Contracts, letters and quotes can be classified by type, urgency or required action — without sending client data to a cloud API. \n
- Decision support. Quotes, proposals or compliance checklists can be scanned and summarised into a yes/no rating, helping you decide quickly without reading every word. \n
Start with exactly one of these. One task, one channel, one clear win.
\n\nWhich model size do you actually need?
\n\nKev ships four sizes together as "Kev 1.0": 0.8B, 4B (the recommended default), 9B, and 27B. On held-out "new source" data, the project's own benchmark puts Kev-27B at 0.851 accuracy against Jev's 0.857 — a one-point gap by their own measurement. Still per the README, Kev-4B on an H100 answers six questions in 18.1ms of model time, and one container serves about 101 requests/second. The 4B model is the one that actually runs locally for many use cases, and it trails the 27B's accuracy by just one point — but if you have no GPU at all, the 0.8B model can still classify simple documents, albeit more slowly.
\n\nThe data rule that keeps you safe
\n\nBefore you paste anything into a tool, one rule matters: treat client data as confidential, even in a friendly chatbot. Under the Swiss Data Protection Act (DSG), do not feed identifiable client information — names, medical notes, financial details — into a public consumer tool without checking. The practical fix is simple: anonymise or replace identifying details first, and for sensitive work in medicine, law or finance, use an approved or local setup. Kev's wire-compatible API means you can run the 4B model on infrastructure you control, keeping your data on premises.
\n\nThat is the honest version of “getting started.” No transformation, no agents, no hype — just one repetitive classification task that no longer eats your afternoon.
