Six Apache 2.0 late-interaction models now cover every commercial retrieval need — mLateOn leads with 57.56 BEIR and nine languages

Seven open-weights ColBERT-style models cleared for commercial use, current to 27 August 2026. Scores are BEIR and MIRACL nDCG@10 as reported on each model card; sizes in millions of parameters. Ranked one to seven by use case, from multilingual flagship to legacy compatibility baseline.

mLateOn is the only shortlist model with a multilingual score — 67.04 MIRACL nDCG@10 across nine explicitly listed languages including Arabic, with 8,192-token documents at 128 dimensions. huggingface.co

mxbai-edge-colbert-v0-32m trades 5 BEIR points for a 10× size cut — 32M parameters and 64-dimensional vectors, yet it handles documents up to 32,000 tokens, the longest context on the list. huggingface.co

LateOn-Code is the only code-search specialist — 149M parameters, 128 dimensions, 2,048-token documents, trained for English and code; no BEIR figure is reported on its card. huggingface.co

ColBERTv2 stays on the list for ecosystem compatibility only — MIT-licensed, roughly 100M parameters, English; it is the legacy baseline much tooling still expects, not a quality pick. thegtmdirectory.com

Read before you deploy

Benchmarks are card-reported. Each nDCG@10 figure comes from its own model card or announcement and may use a different evaluation configuration. Rerun retrieval quality on your own corpus before committing.

Licensing. Apache 2.0 and MIT are permissive licenses generally compatible with commercial use, but this is not legal advice — inspect each repository's terms and dependencies yourself.

Excluded on license. Prominent options jina-colbert-v2 and Reason-ModernColBERT were excluded because their downloadable weights are CC BY-NC 4.0, which bars commercial use.

One correction. The LightonAI ColBERT-Zero row arrived without a computed rank and label due to a string-normalization artifact in the shaping query, not missing evidence; it is restored here at rank 3 as the best compact general model, with the 55.43 BEIR nDCG@10 its primary model card reports.

The shortlist, row by row

#	Model	Best for	License	Params (M)	Dim	Doc tokens	Languages	BEIR	MIRACL	Source

Method: 7 shortlisted open-weights late-interaction retrieval models cleared for commercial use, compiled to 2026-08-27. BEIR and MIRACL columns are nDCG@10 as reported on each model card or announcement, not re-evaluated under a common harness. Parameters in millions; document tokens are maximum indexable document length. Blanks mean the source reports no value. ColBERT-Zero's rank and label were restored from the brief after a shaping artifact; all other figures are unmodified. Noncommercial models (CC BY-NC 4.0 weights) were cut by design.