LightOn releases mLateOn: SOTA multilingual retrieval with just 115M parameters
antoine_chaffin · x · 2026-08-20
LightOn has released mLateOn, a new multilingual ColBERT-style retrieval model that sets a new state-of-the-art among 11 evaluated late-interaction models.
- Performance: With only 115 million active parameters, it delivers retrieval quality comparable to leading 4B and 8B parameter dense embedding models.
- Datasets: The team open-sourced the mLateOn (multilingual) and LateOn (English) datasets used for training.
- Availability: The models and datasets are available on Hugging Face under the "mDenseOn & mLateOn" collection, which also includes related components like DenseOn and LateOn-Code.
More from Research
- Funding Climate Incentivizes Technical Work Over Social Impact — evijit · 2026-08-20
- SPADE enables AI continuous self-improvement via auto-generated environments — pliang279 · 2026-08-20
- dots3-note Preview Shows Adaptation in Unseen Environments — omarsar0 · 2026-08-20
- Critic vs. Environment Rewards: Hidden-Rule Game Case — omarsar0 · 2026-08-20
- TEMPO switches the same model between actor and critic to value long-horizon rollouts — omarsar0 · 2026-08-20
- Reflection on RL: Good for boundaries, bad for long-term goals — sethlazar · 2026-08-20