Scaling Inference: HZDR Joins the Federated Blablador Ecosystem
The demand for large-scale Large Language Model (LLM) inference within the Helmholtz Association has rapidly outgrown localized setups. While FZJ’s Blablador Helmholtz Cloud service provides a solid and rapidly evolving foundation, HZDR sought to push the boundaries of model size and speed with the most modern GPUs.
So, federation it is?
To bridge this gap, HZDR contributed advanced HAICORE GPUs to the fold, establishing a true federated inference service across the entire Helmholtz network. By adding infra power to FZJ’s platform, we can now deploy larger models like the 1-million-token context DeepSeek-V4.1-Flash at unprecedented speeds. The B200 GPUs provide a substantial performance improvement over previous generations, supporting more efficient load distribution and greater stability while meeting applicable technical security standards. At the same time, established models remain available.
This collaboration proves that distributed resources create a powerhouse greater than the sum of their parts.
Where to get it?
The integration is currently running as a pilot: It is already available for individual use – test the service yourself by logging in to the Blablador portal. Simply search for “hzdr” and select “deepseek-v4.1-flash-at-hzdr” to experience the power of our federated setup. API-based access is possible as well.
Joining forces, cool. So… can my center contribute?
Yes!
The technical connection is comparably simple. HZDR’s pilot contribution may serve as a blueprint for other centers wishing to contribute GPU resources and benefit from the Blablador integration. It will be worthwhile to contribute resources and models suited for specific use cases and requirements, and allow seamless access to all of them via the federated platform.
If you have questions, feel free to reach out to us at support@hifis.net.
Oh by the way…
This blog post was formulated with the help of Blablador and Aleph Kolibri, a German model 🇩🇪.
