What actually runs on a laptop, and what it is good for
On-device models are the most over- and under-sold idea in this field at once. Here is the honest version of what a small local model can do.
Following last week’s on-device release, the obvious question: what can a model that runs on your own machine actually do? The answer is narrower than the marketing and broader than the scepticism.
What it does well
Reformatting. Extracting fields from a clean document. Summarising something short. Classifying an incoming message into one of five categories. Rewriting dictated notes into sentences.
These are unglamorous tasks and they are a large share of what a company actually needs. The 16 October argument applies with more force here: most business work does not need the largest model.
What it does badly
Anything requiring broad knowledge it was not given. Anything with several dependent steps. Anything in a language it was not well trained on — and for Swiss companies that is a real constraint, since smaller models degrade faster in German and Italian than large ones do.
Test that specifically before believing any claim about multilingual capability. The method from 14 November takes an afternoon and settles it.
The part people underestimate
It is not just private — it is free at the margin. No per-use billing, no monthly growth, no routing decision you cannot see. For a task that runs ten thousand times a month, that is a different cost structure entirely.
The part people overestimate
That installing it is the project. It is not: the project is everything after. Keeping it updated, connecting it to your documents, and answering the colleague who says the free tool at home gave a better answer — which it will have done, because it is a much larger model.
That expectation gap is the most common reason local deployments get abandoned, and it is manageable if you name it in advance.
Who should look at this in 2026
Companies with a connectivity constraint, a contractual prohibition on external processing, or one very high-volume repetitive task. For everyone else it remains an interesting development rather than a decision.