Artificial intelligence that belongs to you.

Open-weights models are models whose trained parameters are published for download, so nokkela.ai can run them on hardware you control instead of calling someone else's API. That is what makes the on-premise and European-hosted options possible at all: the model file sits next to your data. Licence terms differ from model to model, so we check each one for your intended use and name it in the concept.

nokkela.ai works with a broad catalogue of open-weights models – among them Mistral, Qwen, Gemma 4, GPT-OSS, Nemotron, DeepSeek, GLM, MiMo, MiniMax, Kimi-K2 and Phi – plus speech models such as Whisper for speech-to-text and Kokoro for speech synthesis. The choice is made vendor-independently: use case, the languages involved, the answer quality required and the available GPU memory decide it, and we test the shortlist against your own material before committing.

An on-premise AI system from nokkela.ai needs a GPU server, a local inference service and integration into your existing systems; we specify, procure and install the hardware on your premises. Sizing follows the workload: the size of the chosen model and the number of concurrent users determine how much GPU memory is needed, so we fix that in the concept rather than guessing. Network, storage and backup are planned with it.

An on-premise AI system from nokkela.ai can be operated air-gapped, because the open-weights model and the retrieval index sit on your own hardware and no request goes to an external API. Model files, the inference stack and updates come in through your own change process instead of a live download. The hosted variant cannot do this – it runs in data centres in Europe and needs a connection.

nokkela.ai offers both: hosted means the models run in data centres in Europe, on-premise means they run on your hardware, in your building. Hosted keeps the cost predictable – per token or as a package – and needs no hardware investment from you; on-premise means an upfront investment plus running costs, but the data never leaves your premises and offline operation is possible. Both avoid a third-country transfer to a US cloud.

A RAG knowledge assistant from nokkela.ai first searches your own documents and then has the language model answer on the basis of the passages it retrieved, so staff get answers instead of hit lists. What we need is access to the document sources – file shares, wikis, ticket systems, contracts – plus your access rules, because the assistant must respect the permissions that already exist. We start with one clearly bounded body of documents rather than everything at once.

With nokkela.ai, whatever you put into an on-premise system stays inside your network: the model runs on your own hardware, so prompts, documents and the index built from them are not sent to an external provider and are not available for anyone else's training. In the hosted variant, processing takes place on infrastructure in Europe instead of with a US provider. Where a use case calls for a third-party model, we name it and its terms in the concept beforehand.