Pleias-RAG: Compact Models Prove Small Reasoners Can Ground Answers in Sources

A research team has unveiled Pleias-RAG, a new family of compact reasoning models designed specifically for retrieval-augmented generation (RAG) tasks. The models prioritize a critical but often-overlooked capability: grounding answers in actual source material through direct citations and literal quotes.

The Models and Their Capabilities#

The Pleias-RAG lineup includes two variants: a 350-million-parameter model and a 1-billion-parameter model. Both were trained on synthetic data derived from the Common Corpus, exposing them to multilingual and diverse open-source materials during mid-training. Beyond basic question-answering, the models handle several advanced RAG workflows—including query routing to determine the best retrieval strategy, query reformulation to improve search precision, and source reranking to prioritize the most relevant material.

Performance and Benchmarks#

Testing on standardized RAG benchmarks like HotPotQA and 2wiki shows that both Pleias-RAG variants outperform most models under 4 billion parameters. Remarkably, they are competitive with much larger systems, including Qwen-2.5-7B, Llama-3.1-8B, and Gemma-3-4B. The models are also the first small language models to maintain consistent performance across major European languages while guaranteeing that generated statements are backed by direct source references.

Practical Advantages#

The compact size unlocks immediate benefits for deployment. Smaller models consume less computational power, fitting easily onto resource-constrained infrastructure—an important consideration for organizations without access to large-scale hardware. By building citation and factuality into the model architecture rather than relying on post-hoc verification, Pleias-RAG aims to improve trustworthiness and reduce hallucination risks in generative AI applications.