This paper introduces Blablador, a Large Language Model (LLM) inference server developed by the Helmholtz Foundation and operated at Forschungszentrum Jülich. Blablador is designed to meet the unique needs of the scientific community, offering:

• Privacy-preserving features - Data never stays on servers unnecessarily
• API accessibility - RESTful interfaces for programmatic access
• Custom model support - Flexibility to run specialized research models
• Open-source collaboration - Active contributions to FastChat and the broader AI community

The paper discusses LLM applications across diverse scientific domains:

• Bioinformatics: Protein structure prediction and genomic sequence analysis
• Chemistry: Molecular property prediction and drug discovery
• Climate Science: Climate model interpretation and extreme event prediction
• Astrophysics: Celestial object classification and anomaly detection

The source code is available at HelmholtzAI-FZJ/FastChat under the Apache 2.0 license.

This document uses LaTeX with the IEEEtran conference template.

• TeX Live (or any complete LaTeX distribution)
• biblatex/bibtex for bibliography management

# Build the PDF
make

# Clean build artifacts and restore source files to pristine state
make clean

• make: Compiles the paper using pdflatex (4 passes) and bibtex for proper citation handling
• make clean:
  • Removes all generated files: main.pdf, .aux, .log, .out, .bbl, .blg
  • Uses git checkout to restore main.tex and references.bib to their original state

If you prefer not to use make, the paper can be built with:

pdflatex main.tex
pdflatex main.tex
bibtex main
pdflatex main.tex
pdflatex main.tex

If you use this work, please cite the paper:

@misc{strube2026blablador,
    title={Blablador: A Privacy-Focused LLM Inference Server for Scientific Research},
    author={Strube, Alexandre and Kesselheim, Stefan and Steinbach, Peter and Rushchanskii, Konstantin Z. and von St. Vieth, Benedikt},
    howpublished={\url{https://arxiv.org/abs/}},
    year={2026}
}

Note: This paper is available on arXiv.

• Alexandre Strube - Helmholtz AI, Jülich Supercomputing Centre
• Stefan Kesselheim - Helmholtz AI, Jülich Supercomputing Centre
• Peter Steinbach - Helmholtz AI, Helmholtz-Zentrum Dresden-Rossendorf
• Konstantin Z. Rushchanskii - Helmholtz AI, Jülich Supercomputing Centre
• Benedikt von St. Vieth - Jülich Supercomputing Centre