Private local AI: an LLM on your hardware or on EU GPUs

Local AI is a language model that runs on hardware you control, in your server room or on a rented GPU in the Czech Republic or the EU, so your data doesn't go to US cloud AI providers. Continero, based in Brno, designs, deploys and runs local AI for companies: it picks the model, connects it to documents, email and ERP, and keeps it running. The service lead is David Fogl.

As of 11 Nov 2026, excl. VAT, at Czech retailers: a 128 GB mini PC from about €3,800, a Mac Studio M5 Max with 128 GB about €5,000, NVIDIA DGX Spark and GB10 systems €5,500 to €7,400, an RTX PRO 6000 96 GB card alone about €14,900 and a workstation with one from €21,800. A four-GPU server starts at about €96,600. Prices changed month to month in 2026; the RTX PRO 6000 rose 87% in a year.

As of 11 Nov 2026, excl. VAT: an RTX PRO 6000 96 GB costs $2.49 an hour on RunPod, €999 a month at Hetzner in Helsinki and from €1,040 a month at the Czech data center MasterDC. On the Vast.ai marketplace it starts at $1.40 an hour, but on a third party's machine. RunPod and Vast.ai are US companies, so for sensitive data we recommend a Czech or EU provider.

Usually not on price, but yes on data. Licenses cost about €18 per user per month. In our 3-year model, an own machine with our operations costs about three times as much for 10 people, about the same for 30 people on one GB10 machine, and a comparable amount for 100 people on a two-GPU workstation. Local AI pays off when data can't go to the cloud, when you want AI over your ERP, or for high volumes of similar documents.

In BenCzechMark, the Czech benchmark from Brno University of Technology, Gemma 4 31B leads the models that fit one GPU with a score of 71.8. The best open model of 2026 is DeepSeek-V4-Flash at 81.0, but it needs two 96 GB GPUs or a 256 GB Mac Studio. The benchmark doesn't measure how natural the writing is, so we always test on your own texts.

Ollama and LM Studio are fine for one person and for trying things out. For a company where several people ask at once, we deploy vLLM. In measurements on a DGX Spark, Ollama didn't batch concurrent requests, so eight at once had the same total throughput as one, while vLLM delivered two to four times more.

An open DeepSeek model running on your hardware doesn't send data to its maker. The Czech cybersecurity agency NÚKIB's warning of 9 Jul 2025 explicitly excludes open-source DeepSeek models deployed locally without a connection to DeepSeek's servers. Using the DeepSeek app or API is a different matter: data then goes to China. If you'd rather avoid Chinese models, we deploy Google's Gemma or a Mistral model.

No. In the Artificial Analysis index of November 2026, the best commercial model scores 65 and models that fit one to four GPUs score 4 to 20. Local models are good enough for summaries, email, meeting notes and document search. For long autonomous tasks, complex coding and general knowledge, cloud models are better, which is why we often recommend hybrid.

Yes, through read-only database access. The model asks about orders, stock or deadlines and builds the answer from your data. We check generated queries and work from a table dictionary prepared with your IT. Writing to the ERP stays with a person.

Yes. GDPR applies to personal data wherever the model runs, but running locally removes the transfer to a cloud provider. Under the AI Act, internal assistants are mainly affected by the transparency obligations in force since 2 Aug 2026: people must know they are interacting with AI. Obligations for high-risk systems, such as employee evaluation, were moved to 2 Dec 2027. Leave the legal assessment for your company to your lawyer.

For smaller models and experiments, yes. But the card has only 32 GB of memory and costs about €5,300 to €10,900 excl. VAT in the Czech Republic. That's not enough memory for models that handle Czech and long documents well. For a similar price, a 128 GB machine such as a DGX Spark or Mac Studio makes more sense.

With a free intro call. Then a discovery phase from €1,199 excl. VAT: we model the costs for your data, test models on a sample and recommend hardware. A pilot on one or two tasks has a fixed price you get after discovery and takes 4 to 8 weeks. You buy hardware only once you know the model works on your data.