Search mandate files without sending them.
Internal assistants for recurring questions on contracts, accounts or correspondence. Professional review stays with you.
THE LAB · LOCAL AI WITH OLLAMA
AI on your own hardware – keep sensitive data in the company
For fiduciaries, law firms and SMEs in Liechtenstein and the region: run language models where the files already are. We assess the need, choose hardware and set up Ollama so document search, summaries and internal assistants can be used in the office. Scope is agreed individually.
30 minutes. In person. No obligation. No price and no hardware promise up front.WHO IT IS FOR
Files stay on site.
Questions get answered on site.
Many teams want to use AI without putting drafts and client files into a public chat. Local inference can help – it does not replace access control or professional secrecy.
Drafts and research on your own network.
Summaries, first drafts and search across your documents. Not a substitute for advice, professional rules or confidentiality.
Find knowledge that is already there.
Make manuals, quotes and internal notes searchable locally and prepare recurring texts, with people still signing off.
Not every tool will be local.
Email, backups, updates or a second model in the cloud may still leave the building. We sort that before setup – without a yes/no comparison of vendors.
HOW WE WORK
Needs, hardware, setup.
Not a catalogue product. We start from your documents and from the machines already in the office. NVIDIA DGX Spark is one hardware option, not the default.
- 01
Assess the need
Which tasks should run locally – document search, summaries, internal assistants? Which data must never leave the machine, and which services stay external on purpose?
- 02
Choose hardware
We map existing PCs, workstations and – where the scope fits – an NVIDIA DGX Spark. We do not sell a fixed bill of materials and we do not claim a partnership with NVIDIA or Ollama.
- 03
Set up and hand over
Install Ollama, models and a first workflow. Access, logging, updates and backups are part of the handover; otherwise local inference is just a program on a computer.
Local inference alone does not create data protection or professional secrecy. Access controls, logging, maintenance, backups and any cloud services still in use have to be reviewed as well. For the legal self-check: AI and professional secrecy.
FIRST USES
Search, summarise,
ask internally.
Typical starting points, not promised outcomes. What we build is agreed case by case.
Find documents again.
Index existing PDFs and files locally so the team can ask instead of hunt. Source traces and human approval stay part of the flow.
Summaries and drafts.
Shorten long matters and prepare recurring letters. Nobody sends a draft to clients or authorities without review.
Internal help for the team.
An assistant for internal questions – with a limited corpus, limited rights and traceable use, not a public chat.
HARDWARE
DGX Spark as one option.
Which machine fits depends on model size, number of users and whether the device should sit in the office, a server room or at home. We settle that before anyone buys hardware.
NVIDIA DGX Spark – one possible machine
DGX Spark is a compact NVIDIA workstation for local models. Ollama published measurements for it on 23 October 2025 (firmware 580.95.05, Ollama v0.12.6). That is an external, model-dependent snapshot – not a digihub measurement and not a promise for your install.
digihub.li does not hold this hardware on behalf of clients, does not sell it as a package price and is not a partner of NVIDIA or Ollama.
Other starting points
- existing workstations that are enough for smaller models
- a machine that is reachable only internally
- a different machine later, if the need grows
Which variant makes sense depends on the actual scope – not on a standard parts list.
| Model (Ollama blog) | Size / quantisation | Decode (per source) |
|---|---|---|
| gpt-oss | 20B · MXFP4 | 58.27 |
| llama3.1 | 70B · q4_K_M | 4.423 |
Agreed individually.
Need, hardware, models, documentation and operations are written down after the intro call. There is no public price list, no certification claim and no guarantee against official or professional-body requirements.
No legal opinion, no slogan that replaces the cloud.
We set up the technical side and name open points. Whether a setup meets professional secrecy, data-protection law or internal rules is for your advisers. Local models replace neither backups nor patch discipline.
GOOD TO KNOW
Questions about local AI
with Ollama.
What exactly does digihub set up with Ollama?
Ollama is software that runs language models on hardware you control. We assess the need, help choose hardware, install the software and hand over a first workflow – document search, summaries or internal assistants, as agreed.
Do sensitive data then stay in the company automatically?
No. Local inference can keep prompts on the device. Access controls, logs, maintenance, backups and any cloud services that still run in parallel still have to be assessed. Local inference alone does not create data protection or professional secrecy.
Is an NVIDIA DGX Spark required?
No. It is one hardware option. Whether existing machines, another workstation or a DGX Spark fit depends on the model, the number of users and operations. We make no ownership, partner or price claim about NVIDIA devices.
Who is this for?
Fiduciaries, law firms and SMEs in Liechtenstein and the region that want to search documents, summarise text or use internal assistants without putting every prompt into a public chat. Whether it fits in a given case is what the intro call is for.
What does setup cost?
The intro call is free. Scope and fee are agreed individually. There is no public price list and no package guarantee.
Are the speed figures your measurement?
No. The example decode figures come from the Ollama blog of 23 October 2025 (firmware 580.95.05, Ollama v0.12.6). They depend on the model and are not a current performance promise for your install.
Does this replace the professional-secrecy self-check?
No. Contracts, confidentiality and provider choice remain a separate topic. That is what the AI and professional secrecy self-check is for.
YOUR NEXT STEP
Which tasks should
stay local?
Bring a concrete document or workflow. In 30 minutes we work out whether local models with Ollama fit your business.
Or write to us directlySource of the figures: Ollama, “NVIDIA DGX Spark performance”, 23 October 2025, tests with firmware 580.95.05 and Ollama v0.12.6. Post on ollama.com. digihub.li is not affiliated with Ollama or NVIDIA. This page as of 7 November 2026. Not indexed until launch is approved.