LACONIC bridges the gap: sparse retrieval matches dense model performance while cutting memory use

Dense retrieval has dominated modern information retrieval systems, but the approach comes with a significant cost: substantial memory overhead and heavy reliance on specialized GPU hardware. A new family of learned sparse retrievers called LACONIC shows that sparse methods can match dense model performance while offering practical advantages for real-world deployment.

A more practical approach to retrieval#

Traditional sparse retrieval methods have been overshadowed by dense approaches, which use vector embeddings to find relevant documents. However, sparse retrieval—which leverages inverted indices for efficient search—offers a path to deploying powerful systems on standard CPU infrastructure. The LACONIC family, built on the Llama3 architecture in three sizes (1B, 3B, and 8B parameters), demonstrates this potential.

The training strategy#

The researchers employed a two-phase training framework to adapt causal language models for the bidirectional reasoning needed in retrieval tasks. The first phase uses weakly supervised pre-finetuning to prepare the base model, followed by high-signal finetuning with carefully selected hard negative examples. This curriculum approach proves effective at developing strong retrieval capabilities.

Performance and efficiency gains#

The largest LACONIC variant achieves 60.2 nDCG@10 on the MTEB Retrieval benchmark, ranking 15th on the leaderboard as of February 2026—a result that matches or exceeds many dense approaches. Notably, this performance comes with 74% less memory overhead than equivalent dense models, making it far more practical for systems operating under tight resource constraints. The sparse design allows for efficient search on commodity CPUs without the accelerator costs that dense retrieval typically demands.

Open source and reproducibility#

The team has released both the code and trained model checkpoints, supporting efforts to evaluate and extend sparse retrieval methods in the broader research community. This openness helps democratize access to high-performing information retrieval systems for applications where computational resources are limited.