One fundamental principle always stays the same: the AI model runs locally on the user’s own system. Offline, Hybrid, and Online do not determine where the model runs; they only determine which additional connections Rainbow may use. In Offline mode, Rainbow works without external connections. In Hybrid mode, the model remains local and can selectively use APIs, for example for up-to-date weather information. In Online mode, the model also remains local; in addition, the integrated Rainbow Apex Router can be enabled.

RAG is already part of the private feature set. When needed, users can draw on their own documents and supplied information as an additional knowledge source. If, for example, a PDF, Word file, or another document needs to be evaluated, the relevant content can be used for the current request without retraining the model. The local model remains the foundation; RAG supplements it wherever additional or up-to-date information is needed. At the same time, control over the user’s own data remains a core part of Rainbow. Local data storage is protected with AES-256, with Argon2id as an additional security safeguard. For private users, the key point is simple: the model stays on their own system, personal data can be managed locally, and additional connections are used only where they are actually wanted.

For companies, Rainbow goes significantly further. The Rainbow Enterprise Model is not a conventional business edition and is not simply access to a general AI service. A company receives its own model, individually tailored to that company. The company itself decides which knowledge should become part of this AI. The information selected for this purpose is prepared and deliberately incorporated into the training and adaptation process of the company’s own model. This allows the model to take into account exactly the company knowledge the company itself specifies. The existing German-language and Germany-specific base forms the foundation and is not counted as company tokens. Company-specific knowledge is added on top of that. The finished customer model, together with the company-specific training artifacts and model components created from it, is intended exclusively for that company.

RAG complements the Enterprise Model wherever knowledge needs to be available immediately or for a limited period; it does not replace the model. Company knowledge intended to remain permanent can become part of the company’s own model. New or current information can additionally be brought in through RAG. If, for example, a new Word file, PDF, or report is created today and information from it needs to be used immediately, the Enterprise Model does not have to be retrained. The document can be provided as an additional knowledge source for the current processing task. The distinction is clear: the Enterprise Model contains the knowledge intended to remain part of the company’s own AI, while RAG flexibly supplements it with current documents and information.

The Enterprise Model also runs locally and can be operated within the company’s own infrastructure. The company therefore determines not only which knowledge enters its model, but also which infrastructure it runs on and which external connections are allowed. The existing hardware is reviewed before implementation. Model size and technical configuration are then matched to the actual infrastructure. The result is not an off-the-shelf standard solution, but an environment designed around the technical resources already in place.

Technically, the Enterprise Model encompasses more than the visible interface. It includes model and system architecture, tokenization, preparation of company knowledge, training or model adaptation, RAG, and local inference. The tokenizer splits text into processable units known as tokens. Inference is the actual operation of the finished model – the moment an input is processed and an output is generated. For companies, the decisive advantage lies in how these components work together: dedicated company knowledge, a dedicated model, existing infrastructure, local processing, and controlled external connections. Rainbow’s AES-256 encryption and Argon2id-based protection remain part of the system. Rainbow therefore addresses two clearly different requirements on the same technical foundation: for private users, an organized digital workspace with a local AI model, RAG, and controlled connections; for companies, an additional dedicated model tailored to company knowledge and infrastructure. The principle is the same in both cases: the model stays local. Private users decide which of their own information and external connections are used; companies additionally decide which knowledge becomes part of their own model and which infrastructure it runs on.