✓ Privacy and control over data – Nothing travels to external APIs, so the risk of disclosure or unauthorised use of data is significantly lower. Ideal for confidential documents, legal services, healthcare, public administration, research and development, and manufacturing.
LOCAL AI · ON-PREMISE
Your AI.
Your data.
Under your control.
We help companies set up local AI: we run open-source language models on hardware we tailor to your needs. You can buy the equipment from us or rent it; we install it in any environment or host it in our own data center in Ljubljana.
We prepare an optimised solution tailored to each client — in software and hardware. We build the AI around your data, processes and budget, so you get the most performance for a sensible price.
Your data under your control, in Slovenia.
Open-source models.
Equipment tailored or rented.
Local support.
YOUR DATA STAYS YOURS
What is local AI?
Local AI (on-premise AI) means the artificial intelligence runs on servers you own or on equipment you rent from us — not in a foreign cloud. We use open models (so-called open-weights) that can be downloaded and run on your own hardware; the leading ones are close to commercial solutions in performance, so they are suitable for serious business use.
Your documents, code lists, contracts, e-mail and internal knowledge base stay under your control — in your network or in our data center in Slovenia — and never end up with external API providers.
Your documents
Your network
DC / location of your choice
Why local AI?
✓ Easier GDPR and regulatory compliance – A clear processing location (Slovenia), easier documentation for legal bases, audits and internal protection policies.
✓ Predictable costs – No per-token bills and no surprises at the end of the month. You pay for the equipment (purchase or rental) and support — not for every single prompt.
✓ Tailored adaptability – The model, context length, interface and flows adapt to you: your own knowledge base (RAG), ERP/CRM/DMS integration, tools and automations by agreement.
✓ Independence from vendors – Local AI does not break when an API provider changes prices or terms, and it has no external quotas or vendor limits. It can even run without internet.
✓ Performance that keeps up – The leading open models are close to commercial ones in quality; with advanced optimisations we get the most out of the chosen hardware.
The latest open-source technology
Models (open / open-weights, by choice and licence):
- Qwen (Alibaba)
- DeepSeek
- Mistral
- Google Gemma
- GLM (Z.ai)
- Kimi (Moonshot)
- gpt-oss (OpenAI)
- and others – we choose by task, language and size; several models can run in parallel on the same hardware.
Inference engines:
- vLLM and SGLang – the industry standards for serving large models with high throughput and many concurrent users.
- llama.cpp – a reliable, lightweight engine for GGUF models; excellent for smaller environments and specialised devices.
- Ollama
- TensorRT-LLM
- Strata
- Project Maya
- Colibri Project
- AirLLM
Optimisations that get the most out of the hardware:
- quantisation – we compress the model to a lower numeric precision (e.g. GGUF, AWQ, FP8), so more of the model fits in the same memory and runs faster, with controlled quality loss;
- speculative decoding (e.g. EAGLE-3) – a lightweight helper model proposes the next words in advance and the main model only checks them; generation speeds up by up to 2–3×, depending on the model and load;
- MoE architecture and KV-cache optimisation – with MoE models only part of the model is computed per token, lowering compute requirements and increasing speed; managing the KV cache supports longer context and more concurrent users.
- KV cache – caching of prompt results
We ensure licensing correctness: the model choice takes its licence into account (Apache 2.0, MIT and the licences of individual vendors) and any restrictions for companies in the EU — we advise on this side too.
What local AI can do in practice
Conversation with your own knowledge base (RAG): Documents, procedures, contracts, offers, technical documentation — employees get answers from their own sources with a reference to the source.
Internal AI portal: (Open WebUI or any custom interface) for the whole company, with user roles and access.
Automation and agent flows: Summarisation, sorting, classification, generating offers, customer support, an internal chatbot.
Development tool: Code completion and review for the development team — without sending code outside. (e.g. deepseek Harness)
Sensitive data: Healthcare, finance, legal documents, the public sector — where using cloud AI is legally or contractually restricted.
Also available on request: Speech models, OCR/document processing, translation, image models.
How we get to a solution
Free consultation (60 min) – What problem we are solving, which data and processes we include.
Analysis and design – Choosing the model, context length and load; a hardware plan (GPU / RAM / NVMe). If needed we test the model on your data and in Slovenian.
A tailored offer – Software and hardware; the option to buy or rent the equipment; installation at your premises or hosting in our data center.
Implementation and integration – Installation, optimisation, connection to existing environments, testing with your data.
Support and maintenance – Model and engine updates, monitoring, scaling.
Equipment to your requirements — bought or rented
- Custom configuration: hardware with graphics cards, suitable memory and NVMe storage.
- Rental without a high upfront investment — monthly costs, upgrades by agreement; the same as renting colocation equipment.
- Hosting: at your premises on the local network or in our data center in Ljubljana (redundant power, cooling, DDoS protection, 24/7 monitoring).
Frequently asked questions
Are open models good enough for serious work?
For most business tasks, yes. The leading open models (DeepSeek, Qwen, GLM, Kimi and others) are close in quality to the closed commercial ones, especially when working with your own sources (RAG). Quality in Slovenian depends on the chosen model, so we test it on your examples during the consultation.
How much does local AI cost?
It depends on the model and the scale of use — so we prepare a tailored offer. Costs are one-off (equipment, implementation) or monthly (rental, support), not per token.
What is the difference compared with a cloud API (e.g. ChatGPT)?
Your data stays under your control, there is no dependence on a provider or its quotas, and the model and interface can be freely customised. With constant, heavier load, a local solution can also be cheaper in the long run.
What does renting equipment mean?
We provide the server and GPU hardware, you pay a monthly rental — with no upfront investment, the same as renting colocation equipment.