Local AI means the model runs on a server or computer inside your company, with no data leaving it. It uses open models such as Llama, Qwen or Mistral. It suits sensitive data: contracts, medical records, know-how. The trade-off is less power on the hardest tasks, so a hybrid setup is common.
The three levels of privacy
- Cloud AI under contract: data passes through the provider but, by contract, is not stored or used for training.
- A private EU cloud: a model deployed only for you in a European data centre.
- Local AI: a model on your own server in the office, with no internet connection at all.
When it makes sense
When data is the most valuable thing the company has, or when a contract, the law or a client forbids it leaving: law firms, clinics, accountants, manufacturers with know-how, companies under NDA or public contracts.
What hardware you need
For most business tasks a server or powerful computer with a good graphics card is enough. How powerful depends on the model and how many people use it at once. We work it out during the audit, before you buy anything.
The trade-offs
Open models are a little weaker than the strongest cloud models on the hardest tasks. So we often build a hybrid: sensitive work is processed locally and general tasks go to the cloud.
An example from our work
Krixel runs its AI entirely on the user's computer: no cloud, no per-request fee and no files leaving the machine.