mDenseOn and mLateOn: more signal, less noise for multilingual agentic search

One Retriever, Many Languages, Zero Compromises

mLateOn was trained to retrieve in nine languages plus code. Evaluate it on the full MIRACL benchmark, including 13 languages it never saw in retrieval training, and its average rises above its score on the trained languages alone.

That's the headline result of LightOn's new release: mDenseOn and mLateOn, two fully open 307M-parameter retrievers built on the same open data recipe as DenseOn and LateOn, extended to nine languages, long documents, and code. Two models, trained once, on identical data and one architecture that keeps its edge on languages nobody planned for.

Two Models, Evaluated on Four Axes

mDenseOn and mLateOn are 307M-parameter models trained across nine languages: English, French, German, Spanish, Italian, Portuguese, Swedish, Norwegian and Arabic. LightOn extends the English recipe through translate-train, building a 2.8-billion-pair multilingual corpus, then folds in MIRACL for organic multilingual data, MLDR for long documents, and LateOn-Code's fine-tuning data for code. LightOn evaluates the two models on four axes: English general-domain retrieval (BEIR), multilingual retrieval (MIRACL), long-document multilingual retrieval (MLDR) and code retrieval (MTEB Code).

The result: the same two models hold up on every axis. mLateOn scores 57.56 on BEIR, ahead of every baseline tested, dense or late-interaction, including models twice its size. On MLDR, it hits 87.69 on trained languages and 77.92 across the full benchmark, the best result of any model compared, at any size: token-level matching exploits long documents better than a single pooled vector can. On MIRACL, it posts the best target-language score in the comparison, and on code retrieval it stays within three points of the best model evaluated.

Generalization, proven once already

When we released DenseOn and LateOn, 149M-parameter models that set a new state-of-the-art for their size class on BEIR, we did something uncomfortable with our own results: we decontaminated BEIR against our own training data, stripping out every evaluation example the models could have memorized. The most adversarial setup possible for our own scores. If the results had been riding on benchmark leakage, this is where they would have collapsed.

They didn't: LateOn stayed the top-ranked model. And a sharper pattern emerged across the comparison: every late-interaction model held or improved its rank after decontamination, while nearly every dense model dropped. Token-level matching wasn't memorizing benchmark overlap. It was generalizing past it.

Generalization, proven again, harder

The multilingual results are the same property showing up a second time, on harder ground, and much more strikingly.

mDenseOn and mLateOn share everything: backbone, training data, objectives. The only difference is the representation: one pooled vector per document versus token-level matching. On MLDR, the long-document benchmark, here is what that single difference is worth:

Where the dense model collapses on unseen languages, mLateOn keeps working, at double to triple the score, from the exact same training data. And the hardest cases aren't just unseen languages, they're unseen writing systems: Cyrillic (Russian), Han (Chinese), Devanagari (Hindi), Thai. None of these scripts appears in mLateOn's retrieval training. Its worst language scores about what mDenseOn averages across all thirteen. And it scores higher on Russian, a language it never saw in retrieval training, than mDenseOn scores on any language at all, including the ones it was trained on.

It's not only about unseen languages, either. Arabic sits inside the training set of both models, and the gap there is 36 points. Across the full benchmark, mLateOn takes every one of the thirteen languages, against its dense twin and against every baseline in the comparison.

Translate-train gives dense models a hard edge: strong inside the languages they saw, weak outside them. mLateOn doesn't have that edge. On MIRACL the same pattern reaches even further, to Japanese and Korean: its average actually goes up once the unseen languages are included. There isn't a word of Japanese in its retrieval training data.

It works. Period.

Why this changes the calculus

Building multilingual retrieval usually means paying to collect training data in every language a product might need, or accepting a ceiling the moment a customer shows up speaking something else. Late interaction loosens that trade-off: a representative handful of translated languages is enough to reach strong coverage well beyond them. This is of course not uniform coverage: languages the backbone itself barely saw, like Swahili, still lag. But broad enough to change the default.

Both models run at 307M parameters, small enough for modest hardware and sovereign deployment, no data leaving the environment. Both are fully open: models, the 2.8-billion-pair multilingual corpus, the 16.3-million-sample fine-tuning data across nine languages and code, and the training pipeline.

The pattern holding across two releases

Decontaminated BEIR proved LightOn's architecture holds up outside its training boundaries once. Languages outside its retrieval training show it again.

Pull mLateOn from HuggingFace. Run it on your data. Or test it on the LightOn Console now!

Technical Deep Dive

The full technical write-up is on HuggingFace: training setup, architecture decisions, benchmark methodology, and what we learned along the way.

[Read it here]