Open models versus the cloud: is local good enough?

"Are the open language models good enough?" is the question that decides whether a sovereign AI project happens. The answer is yes for most things, no for some, and the difference is worth understanding before you decide.

Where open models stand

Open models, meaning models whose weights you can download and run yourself, now come from several directions: Qwen from Alibaba, Llama from Meta, Mistral from France, Gemma from Google, DeepSeek from China and gpt-oss from OpenAI. They ship in several sizes, and the best in the 30- to 120-billion class score on benchmarks roughly where the leading cloud models were six to twelve months earlier.

That sounds like a shortfall. In practice it means they handle essentially everything companies actually do with AI today.

Where local is plenty

• Search and questions over your own documents. Here it is not the model's intelligence that decides but how well you feed it the right material. A local model with access to your systems beats a cloud model without it.
• Summaries and drafts. Meeting notes, reports, emails. The difference between models is marginal.
• Classification and structuring. Sorting tickets, extracting fields from invoices, tagging documents. Small models are often better here because they are faster and cheaper to run at volume.
• Everyday coding help. Completion, explanations, tests. Open code models are genuinely good and keep your source code at home.
• Anything involving sensitive data. Here the question is not whether local is enough, but whether the cloud is even an option.

Where the cloud still wins

• The very hardest reasoning tasks. Complex legal analysis, advanced mathematics, long multi-step problems. Frontier models are still ahead, even if the gap shrinks with every release.
• Breadth in rare languages and domains. The largest cloud models have seen more of everything.
• Zero operations. If you do not want to own anything at all and have no sensitive data, an API is the right choice.

Note what is not on the list: Swedish. The leading open models handle Swedish well, and for domain-specific text you can fine-tune on your own documents. Some cloud providers offer tuning too, but then the training data goes to them and the result stays on their platform.

What changes the calculation: fine-tuning

An open model can be trained further on your data. Not from scratch, but a light adaptation that teaches the model your terminology, your templates and your tone. It takes an evening on your own GPU, not a data centre. The result is often a smaller model that beats a larger general one on precisely your tasks. That option only exists when you own the model.

How to decide what applies to you

Do not guess. Test. Take twenty real examples from your business, run them through two or three open models and one cloud model, and have the people who will use the system judge the answers blind. It takes a day and gives an answer that holds for you, not for a generic benchmark.

That is how I do it in the assessment, and it is surprisingly often that the smaller local model wins.