Local LLMs are no substitute for frontier. They are the other half.
There's a loud movement right now for local, open-source or Swiss-hosted language models. We're part of it ourselves — our own anonymisation system deliberately runs on a Swiss-hosted model, for good reasons. Which is exactly why one argument in this movement bothers us as it hardens into dogma.
Open-source models, the claim goes, now lag frontier by only «four months», so local is enough. For an SME in a competitive market, that's the wrong conclusion. And the four months measure the wrong thing.
First, briefly: what does «frontier» mean?
Frontier models are the most capable language models on the market at any given time — the leading edge of what's technically possible today. The best known are Claude (by Anthropic), GPT (OpenAI), Gemini (Google) and Grok (xAI). These models run in the providers' data centres, not on your own hardware. «Local» means the opposite: models you run yourself — on your premises or with a Swiss provider. Freely available, but in practice always a notch below the frontier edge.
Two things get conflated: sovereignty and capability
Data sovereignty is a real argument. Anyone handling sensitive data has good reasons to run a model in Switzerland — we do. But sovereignty says nothing about capability. Where my data sits, and what a model can do, are two different axes. The local camp argues on the first axis and carries the verdict over to the second. This is exactly where a second look pays off.
The «four months» measure the exam, not the practice
The alleged gap is measured on benchmarks. On benchmarks, open models do indeed catch up fast. What doesn't close in four months is what matters in operation: reliable tool use, long context that doesn't fall apart over many pages, and dependability across multi-step processes. A task that has to hold ten dependent steps correctly is something other than a benchmark question. That dimension moves more slowly — and it's the one that carries demanding business processes.
Where the line really runs for an SME
The relevant comparison is not «best local model on a GPU cluster versus frontier». It's «what an SME can run on reasonable hardware versus frontier». On that realistic line there is, today, a capability edge that local can't reach: reviewing a large-firm contract against your internal rules, answering a tender (RFP) cleanly, supporting an ISO certification with all its cross-references, developing software seriously. These aren't repetitive tasks — they're precisely the multi-step, context-rich ones that separate frontier from local.
And here lies the real opportunity: frontier lifts an SME to a level once reserved for large firms with whole specialist departments. A tender for which a corporation assembles a team becomes manageable for the small business. That's not an efficiency gain at the margin — it's competitiveness at enterprise level. Running local only, because it fits the sovereignty narrative, optimises the argument, not the outcome.
Important — and this qualifies us too: for narrowly defined tasks, local can not only suffice but win. In our own system, a compact, Swiss-hosted model beat a far larger competitor on a clearly defined task. Bigger isn't better. But «enough for this one task» is not «enough for the most demanding process in the house».
The answer is a mix — and frontier builds it
In operation we see no either-or, but a division of labour:
- Local / CH-hosted for the repetitive, data-sensitive, high-volume work: classifying, extracting, standard replies — anything where the task is clearly defined and the model is reliable enough.
- Frontier for the demanding work: multi-step reasoning over large, heterogeneous contexts — the contracts, the tenders, the certification, the development.
The crucial point missing from the debate: frontier pays off not only at runtime, but at design time. You use the strong model to build the framework in which the cheap local model runs reliably — the prompts, the checks and test cases, the guardrails. Frontier becomes the foreman for the local fleet. That way an SME gets both: deep capability where it counts, and cheap, sovereign operation for the volume.
Our viewpoint
Local and Swiss-hosted models are right and important — we rely on them ourselves. But they don't replace frontier, they complement it. An SME in a competitive market can't afford to leave the capability edge on the table for ideological reasons.
One point remains: this holds for today. The gap shifts, local models get better, and some of what needs frontier now will run locally tomorrow. Anyone judging the situation seriously dates their verdict — instead of declaring a snapshot a permanent truth. The question stays the same, its answer moves: not «local or frontier», but «which task belongs on which model today — and who builds the framework for it».
Where to start. Before you choose models, you should know the tasks. A sober AI audit does exactly that: it works through the processes in the business and sorts them — what is repetitive and data-sensitive enough for the local model, what is so demanding that it needs frontier, and where the effort isn't worth it at all. That's a more robust basis for an AI strategy than the question of which model is making headlines. Start cleanly here, and you build the mix right from the outset.
Not «local or frontier». The question is: which task belongs on which model — and who builds the framework for it.
Georges Leuenberger in dialogue with Claude Opus 4.8 · As of June 2026