M3-Embedding: Multi-Linguality, Multi-Functionality, Multi-Granularity Text Embeddings

Retrieval systems needed a separate model per retrieval style, language and input length.

One dense model supporting THREE retrieval functions simultaneously - dense, MULTI-VECTOR (late interaction) and SPARSE (lexical) - over 100+ languages and inputs from short sentences to 8192-token documents. Trained with self-knowledge distillation, where relevance scores from the different retrieval functions are combined into a teacher signal for each other.

Reports new state-of-the-art on multilingual, cross-lingual and long-document retrieval benchmarks. No latency, throughput or hardware figures are given, so there is no serving-cost claim to extract.

SLUG RENAMED 2026-10-04 (parent agent): was 'bge-m3', colliding with the models record of the same id, which made every bare-slug cross-reference ambiguous. Inbound refs ([]) all mean the models record, so none needed repointing; papers are cited via `sources`, never by bare slug. Venue ACL Findings 2024 (crossref 'Findings of the Association for Computational Linguistics'). THE SERVING-RELEVANT PART IS THE THREE-FUNCTION SUPPORT, and it cuts against the single-vector orthodoxy: this one checkpoint can emit a dense vector for ANN search, a sparse vector for inverted-index retrieval, and multi-vector ColBERT-style vectors for reranking. That collapses three separate model deployments into one, and the sparse head in particular is the cheapest of the three at query time. The multi-vector head carries the ColBERT memory cost documented in pap-mm-colbert. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'findings'. Located in the ACL Anthology bib index: anthology id 2024.findings-acl.137, booktitle 'Findings of the Association for Computational Linguistics: ACL 2024'. The record's `venue` stays 'acl' - which is not wrong, only incomplete - but the track is now explicit. Title note: the anthology title is 'M3-Embedding: Multi-Linguality, Multi-Functionality, Multi-Granularity Text Embeddings through Self Knowledge Distillation'; the record name drops the trailing subtitle, which is why a first exact-title match missed it and a fuzzy match was needed. No status change: `venue` is untouched.