LightOn Releases Fully Open Multilingual Retrieval Models mDenseOn & mLateOn
antoine_chaffin · x · 2026-07-30
LightOn has released mDenseOn and mLateOn, two fully open 307M-parameter multilingual retrieval models.
- Data & Training: They extended their validated English recipe to 8 additional languages via translate-train, creating a massive 2.8B-pair pre-training corpus and 16.3M fine-tuning samples covering natural languages and code.
- Performance: The models achieve SOTA on benchmarks like BEIR (English), MIRACL (multilingual), and MLDR (long-context). Notably, multilinguality does not degrade English performance, with mLateOn hitting 57.56 on BEIR.
- Generalization: The late-interaction architecture (mLateOn) shows superior zero-shot transfer to unseen languages, scoring 67.59 vs 57.42 for mDenseOn on the MIRACL unseen-language split.
Models, datasets, and training code are fully open-sourced.
Related event: LightOn Open-Sources Multilingual Retrieval Models mDenseOn and mLateOn(18 posts)→
More from Models
- Rumor: GLM 5.5 and Kimi K3.1 dropping soon, possibly slowed by anti-distillation blocks — zephyr_z9 · 2026-10-16
- Thrive Holdings benchmarks AI on real tax and IT work across 75+ firms — BorisMPower · 2026-10-16
- A Little Encouragement Makes a Small Local Model Draw a Stunning Pelican SVG — Fancy-Snow7 · 2026-10-16
- Rapid-MLX 0.16.2 brings NVIDIA Nemotron 3.5 up to 120B to Apple Silicon Macs — kuchaev · 2026-10-16
- "End of Chinese OSS" claim pushed back: open models get license almost automatically — eliebakouch · 2026-10-16
- Opus 5.5 still leads a month later, but open-weight decision models are closing in — airesearch12 · 2026-10-16