Secure AI in your company?

What local models are

The best-known AI models run in the provider's cloud - every query is sent to their servers and processed there. A local model is deployed on your own infrastructure - a computer or a company server - and all of the processing happens inside your organization.

There is no per-seat subscription and no charge per query, and the content of the conversations - contracts, documents, correspondence - never leaves the company.

That matters wherever trade secrets or, say, contractual obligations to business partners are at stake. The system can also run on a network cut off from the internet.

Sample hardware requirements

Performance depends directly on the configuration: memory bandwidth and GPU power determine how many tokens are generated per second, while the amount of memory determines how large a model can be loaded. A model will run on a weaker machine too, but at the cost of response time.

Extension example - web search

Out of the box, models of this kind have no access to the internet. Their knowledge comes solely from training data, which covers material up to a fixed cut-off date - usually a year or more before the model is released. Questions about current events, today's exchange rates or recent changes in the law lead to answers that are out of date or simply made up.

In our demo, the "Web search" toggle adds a search layer on top of the model: the question goes to a search engine first, and the model answers based on the results it gets back, citing its sources. This is an extension built by Kreatio, not a feature of the model itself - without it, the model knows nothing beyond its training data.

Extension example - report mode

Here the model works as an agent: it composes the search queries itself, picks the sources worth reading from the results, opens them in a real browser and only then writes the report from their content - with footnotes you can check.

It is a simplified version of what OpenAI calls deep research and Anthropic calls Research. In our demo the order of the steps is set by code, there are just a few sources, and the report is ready within a minute. The principle is the same: hours of searching turned into minutes, and an answer based on what the model has actually read, not on what it remembers from its training.

Same model, different result

Training models is the domain of a handful of labs worldwide, at a cost running into hundreds of millions of dollars. But the value a model brings to your company comes from something else: the tools, the data and the way the task is orchestrated around it.

Claude Code is the best example. In a chat window, the same model does nothing but answer questions. Give it tools, access to files and an action loop, and it writes code on its own and fixes its own mistakes. The model has not changed. What surrounds it has.

We can design, build and tailor all of this to your needs: search across your company's knowledge base - answers grounded in your documents; tool calling - reaching into your systems and data; agents - carrying out multi-step tasks from start to finish.

All of it locally and securely, even with no internet access at all.

Pros and cons

Try the model