The three places
- Where the files are stored. This is the easiest and most common promise: “We don't store anything in the cloud.” It says nothing about what was sent in the meantime.
- Where the program runs. A program on your computer instead of a browser window. Convenient, and almost irrelevant for confidentiality.
- Where the model computes. This is where it is decided. Everything the model has to see in order to answer goes to wherever the model is.
That is why “no cloud storage” is not the same as local processing. The statement is true, but it answers a different question. Anyone who offers it as an answer to the confidentiality question is dodging, sometimes without realising it.
Four operating modes: what leaves your computer in each
| Level 1 · pure web service | Files, history, inputsUploaded content and the history are processed by the service and stored there. The scope depends on the product, the settings and any connected services. |
|---|---|
| Level 2 · program on your side, model at the provider | Task text, context, file excerptsFiles and tools stay on your computer; whatever the model needs for the task goes to the provider. Where it computes and who can access it there decides the rest. |
| Level 3 · model at a provider in the EEA, without retention | The task text, processed in the EEAThe third-country question eases, the access question remains, and retention commitments are usually secured by contract, not by technology. For professional secrecy, this is not the same as level 4. |
| Biwak | Between levels 2 and 3Files, conversations and checkpoints stay on your computer. Your question, the context and the file contents used go via the Biwak account in Frankfurt am Main to Microsoft Azure in the EU Data Zone; the contracting party is Microsoft Ireland Operations Limited. If Azure rejects a task, it goes to PREM SA in Lugano, Switzerland, and from there to its compute partners Nebius and TensorX in the EU and the United Kingdom. Both providers rule out training. Microsoft may store suspicious requests for abuse detection and have them reviewed by humans; Prem rules out storage by contract, without technical safeguards. New model tasks require an internet connection. The route in detail. |
| Level 4 · model in your own network | No external model request neededOther tools, web search, telemetry and remote maintenance need to be checked separately; a local model alone does not guarantee an isolated workflow. Biwak does not offer this operating mode. |
What the supervisory authorities say
“Technically closed systems are therefore preferable from a data protection perspective.”
The guidance on technical and organisational measures for AI systems of June 2025 recommends deliberate decisions about local use and necessary online connections. It calls local processing on the end device “preferable” for retraining or fine-tuning the model. At the same time, it points out that when a model runs on end devices, the personal data it needs may have to be distributed to all devices, which weights the risks differently. Both papers are positions of the supervisory authorities, not law and not a court ruling.
Our assessment: a model in your own network is the strongest route for confidentiality. But it requires computing power, maintenance and expertise that smaller offices often lack. The next section explains what that means in practice.
What your own model costs
If you want to run a model yourself, you need to check three things separately:
- Hardware. What you need depends on the model version, quantisation, context size and the speed you want. Check the available memory and measure with your own sample files; a blanket memory figure does not prove suitability.
- Model quality. Test the same tasks with the same acceptance criteria as for your current tool. Summarising, analysing figures and reproducing sources are different requirements; the label “local” says nothing about the quality of the results.
- Operation. Updating models and programs, managing access, backups and troubleshooting all take effort. Measure it in your own pilot instead of assuming a generic maintenance time.
Programs for local models are also worth a second look. Besides local models, Ollama now also offers cloud models; if you want to be sure that computation stays local, the Ollama FAQ says to switch off the cloud features. According to its own documentation, LM Studio works without internet once the model files have been downloaded. For model tasks, Biwak uses the managed Biwak access; you cannot enter a model server of your own. How Ollama and LM Studio compare is shown in the comparison with Biwak.
How to check it instead of taking it on trust
- Test without a network. After setup, run a new sample task without a network connection. This shows whether this workflow works locally, but it does not prove that nothing is transmitted during online operation. With Biwak, new model tasks require a connection.
- Check the connections. Separate the model requests from other network access by the tools, such as fetching web pages or updates.
- Check the processing chain. Who computes, in which country, with which sub-processors, and what do the contracts say about retention and training?
- Ask about remote maintenance. Screen sharing and content-related telemetry create exactly the kind of access that the word “local” is supposed to rule out.
- Ask about the log. If it is kept on your side, you can check it yourself when in doubt. If it is kept by the provider, you have to ask, and the answer is then evidence about evidence.
File access on your device and account and usage data at the provider need to be considered separately. Biwak needs the latter for access, quotas and billing; exactly what is collected is set out in the privacy policy.
Changes to this article
- 4 October 2026: Biwak row and answers on the model route brought up to date: model tasks now run primarily at Microsoft Azure in the EU Data Zone, and via PREM SA only as a fallback.
- 24 September 2026: Added a question on free use of LM Studio and Ollama in companies, supported by the vendors' blog and licence; merged two questions about Biwak.
- 23 September 2026: Biwak added as a separate row in the table of operating modes, with the chain via PREM SA; the statement on the June 2025 guidance narrowed to retraining and fine-tuning; title no longer contains the word “offline”; removed the reference to a specialist article whose wording could not be retrieved during review; added a note on Ollama's cloud models.
- 10 September 2026: Narrowed the statements on complete isolation, hardware, model quality and maintenance effort.
Frequently asked questions
Is a tool “local” if it does not upload my files?
Only in a weak sense. If the task text, together with excerpts from your files, goes to someone else's model, the content is out, even if the file stayed where it was. With a draft contract or a patent application, the content is precisely the secret.
What hardware do I need for a local model?
That depends on the model, its version, the input size and the speed you want. Check the model's requirements and test with representative tasks on the intended hardware. A general figure in gigabytes is no proof of good results.
May a company use LM Studio or Ollama free of charge?
According to the vendors, yes. Since 8 July 2025, LM Studio has also been free for use at work; before that, its terms of use required a separate licence for this. Ollama is under the MIT licence, and local use is included in the Free plan. Check the licence of each model separately.
Can Biwak work without internet?
Not for new model tasks: your question, the context and the file contents used go via the Biwak account to Microsoft Azure in the EU Data Zone, or as a fallback via PREM SA in Switzerland to computing partners in the EU. Locally stored files and checkpoints stay on your device. If you need to work entirely offline, you need a model on your own hardware.
Sources
- DSK: Guidance on artificial intelligence and data protection, 6 May 2024 (PDF)
- DSK: Guidance on technical and organisational measures for AI systems, June 2025 (PDF)
- Ollama: FAQ (local operation, cloud models, switching off cloud features)
- LM Studio: Offline Operation
- LM Studio Blog: LM Studio is free for use at work, 8 July 2025
- Ollama: licence (MIT) in the source code repository
- Biwak: Privacy policy, section 14 (requests to a language model)
This text is not legal advice. It is the groundwork we had to do for ourselves, with the legal references, so that your lawyer does not have to start from scratch. Where a question depends on your circumstances, the text says so.