Private AI for companies
Private AI is a language model that runs on your own server or in a data centre you choose, so documents and customer data never reach outside clouds. It suits clinics, financial firms and anyone handling biometric data, which Uzbek law requires to stay in the country. Projects start at $15,000 and take eight weeks or more.
Cloud models are convenient, but every request goes to the provider's servers. Often that's fine. Sometimes it isn't: since the March 2026 amendments, Uzbek law still requires biometric, genetic and telecom subscriber data to be stored and processed inside the country. Contracts, security policies and offline sites add their own constraints.
When you need it
Signs the cloud isn’t enough
- you process biometric, medical or genetic data
- you are a telecom operator or handle subscriber data
- client contracts prohibit transfer to third parties
- your security team won’t approve external APIs
- the system has to run without internet access
Which model
Open models such as Qwen, Llama, Gemma, Mistral and DeepSeek now match cloud models on many applied tasks: document extraction, answering from a knowledge base, classifying requests. They still trail on long reasoning, and Uzbek output quality varies, so we benchmark candidates on your own examples before anyone buys hardware.
| Model size | GPU memory | Good for |
|---|---|---|
| 7–8B parameters | about 6–8 GB | classifying requests, extracting fields |
| 30–32B | about 20–24 GB | knowledge base answers, customer replies, drafts |
| 70B | about 48 GB or more | complex answers, long documents, several departments |
Rough figures for 4-bit quantisation from public guides; long context and many concurrent users need more
What the project includes
Model selection with quality benchmarks on your tasks, server sizing and purchase or rental, deployment (usually vLLM for production), a knowledge base and agents, access control and request logs, and training for your administrator. From $15,000 excluding hardware, eight weeks or more. A hybrid setup, where sensitive data stays local and everything else uses a cloud model, is often the cheapest option. Legal background: Personal data law in Uzbekistan and AI.
Is a local model worse than ChatGPT?
On narrow tasks such as document extraction the gap is often small; on complex reasoning cloud models are stronger. We compare both on your data first.
Can the model run in a data centre in Uzbekistan?
Yes, which satisfies the localisation requirement without your own server room.
Who keeps it updated?
We do under a support plan from $390 a month, or your administrator after training.
