IPCEI-CIS Hackathon @ SAP Innovation Center Potsdam
In July 2026, the Thalamus team joined a hackathon at the SAP Innovation Center in Potsdam, collaborating with contributors from the NeoNephos Foundation and ApeiroRA as part of the broader IPCEI-CIS initiative to develop open-source European cloud and AI infrastructure.
During the event, we worked on three concrete topics. We built a Naira plugin that feeds Thalamus model data into Naira's catalog, making it possible for Naira's MCP server to answer live questions about running inference instances. We also worked on two Open Component Model (OCM) topics: packaging Thalamus itself as an OCM component and distributing model weights through OCM using a Hugging Face-compatible API.
Naira Integration
We built a Thalamus plugin for Naira that keeps Naira's AI asset catalog in sync with the models currently running in Thalamus. With the catalog populated, Naira's MCP (Model Context Protocol) server lets an MCP-capable client such as Open WebUI ask live questions about Thalamus-managed inference instances. For example, clients can ask which models are available and what engine configuration they use.
OCM: Packaging Thalamus as a Component
Thalamus today is installed through a set of Helm charts and configuration files bundles through a helmfile. Together with the OCM team, we explored packaging the whole Thalamus installation as a single OCM component. The goal is to turn Thalamus into a versioned, transportable artifact that can be pulled from a sovereign OCI registry and installed without relying on external tooling or manual setup. This work is still experimental and needs to mature before it can replace the current installation path.
OCM: Model Weights as an OCM Component
Thalamus currently relies on external registries such as Hugging Face for model weights, which creates a dependency outside our control for a sovereign-cloud offering. We prototyped an integration with the OCM model-server to store model weights as OCM components in an OCI registry of our choice while still exposing them through a Hugging Face-compatible API. During the hackathon we packaged a lightweight demo model, stored it in GHCR, pulled it through the model-server, and served inference from it without the weights ever touching Hugging Face.