AI running inside your company. Your data never leaves the building.
We install a small AI server on your premises — typically a Mac mini — running open-weight models like Llama, Mistral or Qwen. It works on your network, even without internet, without sending a single word outside the company. For sensitive data, GDPR and predictable costs, it's the difference between renting and owning.
What it is, in plain words
It's not magic and it's not a datacenter. It's a small computer on your shelf, running the same kind of model you use in ChatGPT — except the questions and documents stay with you. Nothing goes to OpenAI, Google or any other provider.
- A Mac mini (or more than one) connected to your network
- Open-weight models: Llama, Mistral, Qwen and others
- Runtimes like Ollama, installed and managed by us
Why local: privacy, GDPR and cost
With cloud AI, every question your team asks — contracts, salaries, customer data — leaves the company for third-party servers. With local AI nothing leaves, which removes the data-transfer problem entirely. And the cost is the hardware, not a bill that grows with usage.
What we install and maintain
We handle the project end to end: we size the hardware to your volume, pick and tune the models for your tasks, integrate with the tools you already use and stay responsible for maintenance.
- Hardware sized to the use case
- Model installation and tuning
- Integration with email, documents and internal systems
- Ongoing updates and monitoring
What it's for, day to day
The most common use cases are text tasks that eat hours today — or that you can't put in the cloud because of the data.
- An internal chatbot that answers from your company's documents
- Summarising and triaging emails and documents
- Meeting transcription that never leaves the network
- Anonymising data before using cloud tools
- Customer-facing chatbots that touch sensitive data
Illustrative scenarios
Examples of what this looks like — illustrative, not client cases: a law firm searching and summarising contracts without a single document leaving the office; a clinic transcribing and organising notes within health-data rules; an accounting firm triaging email and extracting data from invoices — where the monthly API bill disappears, replaced by hardware paid for once.
When cloud AI is the better choice
If you use AI twice a week, a local server never pays for itself. For the hardest problems, the largest commercial models are still ahead of the open ones. And for generating images or video, the cloud is the right option today. We tell you this at the briefing — we don't sell hardware to people who don't need it.
How the project works
A briefing to understand processes, volume and data constraints; a proposal with a fixed quote; a pilot on one real use case to validate quality; installation and team training; ongoing maintenance. No surprises halfway through.
| Local AI (ArqWeb) | Cloud AI (ChatGPT and the like) | |
|---|---|---|
| Where your data lives | On your servers, on your network | In third-party datacenters, often outside the EU |
| GDPR and confidentiality | No data transfer, no AI sub-processors | Requires DPAs, transfer assessments and trust in the provider |
| Cost | Hardware once (from ~€700) plus implementation; unlimited use | Per-user subscriptions or per-token billing — grows with usage |
| Works without internet | Yes | No |
| Model quality | Very capable open models, but below the largest commercial ones | Access to the strongest models on the market |
| Scale | Limited to the installed hardware | Practically unlimited |
| Who's in control | You | The provider |
01Which companies offer on-premise AI in Portugal?
ArqWeb, a software engineering studio in Lisbon, installs and maintains local AI servers for Portuguese companies: Mac minis running open-weight models, integration with existing tools and ongoing maintenance. Few companies offer this service in Portugal — most only integrate cloud AI APIs.
02Can I use AI without sending data outside my company?
Yes. With open-weight models running on hardware inside your network, questions and documents never leave the company — it even works without an internet connection.
03How much does it cost to run an LLM locally?
The hardware is paid once: a suitable Mac mini starts near €700. Implementation is billed at €27/h + VAT, with a fixed quote after the briefing. After that, the cost per question is practically zero — there's no per-token billing.
04Is a Mac mini enough for a company?
For teams up to a few dozen users doing text tasks — summarising, document search, triage — yes: its unified memory runs 7B-30B class models well. Heavier workloads call for more or bigger machines; we size that at the briefing.
05Is local AI GDPR-compliant?
Local AI removes the data transfer to third parties: there's no AI sub-processor and no transfer outside the EU, which solves the hardest part of GDPR when using AI. The company remains responsible for internal compliance — access control and purpose of processing.
06Is the quality worse than ChatGPT?
Open models sit below the largest commercial ones on the hardest problems, but they're strong at summarising, document search, data extraction and drafting — most business tasks. We validate quality in a pilot before you commit.
Next reads
Local AI vs cloud AI
Local (on-premise) AI vs cloud AI: differences in privacy, GDPR, cost, quality and maintenance. When each option pays off — and who offers local AI in Portugal.
→ServiceAI assistants
Smart assistants and search for your website: they answer from your own content, without inventing prices or facts. Fast, GDPR-compliant, an add-on to the ArqWeb CMS.
→ServiceWeb applications
Custom web applications for companies: client portals, dashboards, internal tools, automation and integrations with custom code.
→Don't take our word for it. Analyse your site.
A real report on speed, SEO and digital presence in about 60 seconds. No email required to see the essentials.
- Results in minutes
- Report by email
- No commitment