The model
Qwen3.8-27B: an open-licence model with 27 billion parameters, used as published. The weights are a file on the server, not a subscription. A small open model (Qwen3-Embedding-0.6B) builds the memory, on the processor.
Kibo / Facts
The model, the server it needs, the systems it connects to and where it runs today. What is not measured or not certified yet is said so.
Qwen3.8-27B: an open-licence model with 27 billion parameters, used as published. The weights are a file on the server, not a subscription. A small open model (Qwen3-Embedding-0.6B) builds the memory, on the processor.
On your own servers, or hosted for you on our servers in Europe. The platform, the model and your data sit on the same server; nothing that reads or writes your data runs anywhere else.
One data-centre GPU with 80 GB of memory (H100 class) for the model, and 8 processor cores with 8 GB of memory for the platform. Smaller machines are being measured; they are named here once they are.
Mailboxes on Microsoft 365 and Google Workspace, shared files, Microsoft Teams and Slack, and company sign-in. Any other system through a standard: an MCP server, or an API with an OpenAPI description. Kibo acts with each person's own rights there.
In pilot at TEKBH and EBH, on their own work, with their own people.
Kibo has no security or AI certification of its own yet. The route to one is being prepared; this page will name it.
Accuracy, how often a person corrects what Kibo proposes, and response times are measured on the pilots. They are published here once they are good enough to stand behind.
More: sovereignty · security · what is inside