RAGU releases 7B GraphRAG model that tops Qwen2.5-32B by 12.5%

A quiet arXiv drop this month makes a claim worth pausing on. A 7 billion parameter model, purpose-built for extraction, reportedly beats a 32 billion parameter general-purpose model on the specific job of turning documents into knowledge graphs. Not by a token, by 12.5% relative harmonic mean, according to the RAGU paper posted by Mikhail Komarov and colleagues.

RAGU itself is a modular GraphRAG engine that separates extraction from consolidation, running entities and relations through DBSCAN-backed deduplication, LLM summarization, and Leiden community detection. The specialized model, Meno-Lite-0.1, is released on Hugging Face under Apache 2.0, while the engine is MIT-licensed and installs with pip install graph_ragu. It runs on a single GPU.

The interesting claim underneath the benchmark numbers is a hypothesis about scaling. The authors argue that the language skills a RAG pipeline actually uses, reading a passage, pulling entities, following a relation, grow weakly with model size, while world knowledge scales steeply. If that framing holds, the field has been overpaying for extraction by reaching for frontier-scale models where a small domain-adapted one would do. Their own reported numbers on GraphRAG-Bench Medical, evidence recall up to 0.84 versus 0.76 or lower for competitors and overtaking HippoRAG 2 on synthesis tasks, are the evidence they offer.

The honest caveat is that this is a single paper, self-reported, and the strongest results are shown on a medical benchmark and on English GraphRAG tasks. What the paper does not give you is independent replication, production latency at enterprise scale, or performance on the messier non-English domains most companies actually operate in. Take the harmonic-mean win as reported, not settled.

Even hedged, the direction is what matters. If small specialized models on open pipelines can carry the extraction stage of GraphRAG, on-prem deployments become feasible for teams that could not justify frontier-scale compute, and the MIT license means the negotiation is with your own infra rather than with a vendor.