SPLADE-Code brings learned sparse retrieval to code search at up to 8B parameters

October 24, 2026, 16:38Source: NAVER LABS Europe

Researchers at NAVER LABS Europe have introduced SPLADE-Code, which they describe as the first large-scale family of learned sparse retrieval models specialized for code retrieval. According to the authors, the 600M-parameter class reaches state-of-the-art results among retrievers under 1B parameters, with a score of 75.4 on MTEB Code.

Key points

NAVER LABS Europe introduces SPLADE-Code, described as the first large-scale family of learned sparse retrieval models for code (600M-8B parameters).
The models score 75.4 on MTEB Code among retrievers under 1B parameters, and 79.0 with 8B, according to the authors.
Training uses a lightweight one-stage pipeline.
Learned expansion tokens are critical to bridge lexical and semantic matching.
Sub-millisecond retrieval on a 1M-passage collection is reported with little effectiveness loss.

Why sparse retrieval for code is hard

Retrieval over large codebases is a key component of modern LLM-based software engineering systems, the paper says. Most existing approaches rely on dense embedding models, while learned sparse retrieval (LSR) has remained largely unexplored for code.

The authors, Simon Lupart, Maxime Louis, Thibault Formal, Hervé Déjean and Stéphane Clinchant, list several reasons why the approach is difficult to apply to code:

subword fragmentation
semantic gaps between natural-language queries and code
the diversity of programming languages and sub-tasks
the length of code documents, which can harm sparsity and latency

The work is published in the Findings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), held in Budapest, Hungary, from 24 to 29 October 2026.

What SPLADE-Code delivers

SPLADE-Code is a family of models ranging from 600M to 8B parameters. The authors say it is trained with a lightweight one-stage pipeline. Despite that simplicity, they report state-of-the-art performance among retrievers under 1B parameters, with 75.4 on MTEB Code. At larger scale, the 8B model reaches 79.0, which the authors describe as competitive results.

The paper also reports that learned expansion tokens are critical to bridge lexical and semantic matching. In other words, the models add useful terms beyond those literally present in a query or code passage, which helps connect natural-language questions to code.

Latency and what it means

The authors provide a latency analysis showing that LSR enables sub-millisecond retrieval on a 1M-passage collection with little effectiveness loss. This addresses a practical concern for systems that must search large codebases quickly.

For teams building LLM-based software engineering tools, the results suggest that sparse retrieval is a viable alternative to the dense embedding models that currently dominate the field. The source does not give details on release or availability of the models, so those remain to be confirmed from the paper itself.