LACONIC Sparse Retrieval Models Match Dense Performance at 74% Less Memory, Paper Shows

Researchers have introduced LACONIC, a family of learned sparse retrievers based on the Llama3 architecture that achieves competitive performance against dense retrieval models while using 74% less index memory. The paper, published on arXiv, demonstrates that the 8-billion-parameter variant reaches a state-of-the-art 60.2 nDCG@10 on the MTEB Retrieval benchmark as of February 5, 2026, ranking 15th on the leaderboard while operating on commodity CPU hardware.

The work addresses a persistent tension in information retrieval: dense models deliver high accuracy but require GPU acceleration and substantial memory for vector similarity search at scale. Learned sparse retrieval offers an alternative by using inverted indices, which are more memory-efficient and run on standard CPUs, but has historically lagged behind dense approaches in effectiveness.

Key facts
| Metric | LACONIC 8B Value | Context |
|—|—|—|
| MTEB Retrieval nDCG@10 | 60.2 | 15th on leaderboard as of Feb 5, 2026 |
| Index memory reduction | 74% | Compared to equivalent dense model |
| Model sizes released | 1B, 3B, 8B | All based on Llama3 architecture |
| Deployment hardware | Commodity CPU | No GPU required for search |

Two-phase training curriculum

The LACONIC team designed a streamlined two-phase training process to adapt causal language models for retrieval. The first phase, weakly supervised pre-finetuning, converts causal LLMs into bidirectional encoders that can understand context from both directions of a query or document. This step uses large-scale weak supervision to establish basic retrieval capabilities.

The second phase, high-signal finetuning, uses curated hard negatives — non-relevant documents that closely resemble relevant ones — to sharpen the model’s discriminative ability. This approach allows the model to learn subtle distinctions between highly similar pieces of text, which is critical for high-quality retrieval in production systems.

Performance on MTEB Retrieval benchmark

LACONIC’s 8B variant achieved 60.2 nDCG@10 on the MTEB Retrieval benchmark, a widely used evaluation suite for retrieval models. This places it 15th on the leaderboard, directly competing with dense models that typically dominate the top ranks. The 1B and 3B variants offer smaller, faster alternatives for applications with tighter compute budgets.

The benchmark results are notable because they show that sparse retrieval can now approach dense model quality without the associated hardware requirements. For organizations running search at scale, this could reduce both infrastructure costs and latency.

Memory efficiency and practical deployment

The 74% reduction in index memory compared to an equivalent dense model means that LACONIC can handle larger document collections on existing hardware. Dense models store embeddings as high-dimensional vectors, which scale linearly with the number of documents and require GPU-based approximate nearest neighbor search. Sparse models like LACONIC use inverted indices, which compress storage and allow exact scoring on CPUs.

This makes LACONIC particularly relevant for real-world search applications where GPU resources are limited or expensive. The ability to run on commodity CPU hardware with a fraction of the compute budget could lower the barrier to entry for smaller teams and organizations deploying search at scale.

Open source release and reproducibility

The authors have fully open-sourced the code implementation and trained model checkpoints. This allows the research community to reproduce results, build on the work, and integrate LACONIC into existing retrieval pipelines. The release includes the 1B, 3B, and 8B variants, giving users flexibility based on their performance and resource requirements.

The commitment to open source aligns with broader trends in AI research toward transparency and reproducibility. For developers and engineers evaluating retrieval models, having access to trained weights and training code accelerates integration and customization.

Limitations and next checks

The paper does not provide detailed latency measurements for CPU-based inference at production scale, which would be helpful for teams evaluating deployment costs. The MTEB leaderboard position is as of February 5, 2026, and may have changed since publication. Additionally, the effectiveness of the 1B and 3B variants on specific retrieval tasks beyond the benchmark remains to be explored by independent evaluators.

Researchers and practitioners interested in testing LACONIC can access the code and checkpoints through the project’s GitHub repository. Further evaluation on domain-specific datasets and comparison with other recent sparse retrieval methods would help clarify where LACONIC fits in the current retrieval landscape.

Source: arXiv cs.CL, “LACONIC: Dense-Level Effectiveness for Scalable Sparse Retrieval via a Two-Phase Training Curriculum,” https://arxiv.org/abs/2601.01684

Datos clave