Markus Fischer, Jens Meiler and Matthias Elgeti build Munin, a self-hosted assistant that searches a research group’s papers, reads full passages and cites its evidence. Its main contribution is the surrounding software and tests, rather than a new model.
Four specialized agents handle reading, search, calculation and longer research reports. The Apache-2.0 release includes document ingestion, a citation graph and tests for missing sources. Groups can rerun those tests after changing their model or collection.
What the authors measured
- On 199 literature questions with source papers present, Qwen3.8-27B answers 87.4% correctly with the full system, versus 38.7% unaided and 21.1% with fixed top-five retrieval.
- That 87.4% result allows web access. Blocking outbound requests lowers accuracy to 66.3%, separating local-document benefits from outside help.
- After removing source papers, the same model correctly abstains on 88.9% of 27 questions it previously answered correctly. The denominator is a small, conditional subset.
- The reference deployment uses two RTX 5090 cards with 32 GB each and a 68,500-paper collection. Groups share access to that corpus; group labels filter relevance rather than enforce isolation.
Caveat: the licensed reference corpus is not released, and a fresh installation starts empty. Public scorecards permit checking the statistics, but recreating the experiment requires assembling the relevant papers.
The privacy boundary also needs attention. Models and indexes run locally, but the reference frontend uses a rented host and web search sends queries outside. The paper reports results from one consortium, with different sampling settings between baseline and agent runs.
The useful adoption pattern is to test failures deliberately: withhold a source, block the network, swap the model and inspect what changed. Its evaluation suite is a stronger starting point than assuming that any compatible model inherits the headline score.