mLateOn tops multilingual ColBERT retrieval with 115M active params, rivals 8B dense models
IgorCarron · x · 2026-08-20
LightOn's mLateOn (115M active / 312M total params) leads multilingual ColBERT-style retrieval on HAKARI-Bench: 65.52 Overall Macro, 8.33 points ahead of runner-up pplx-embed-v1-late-0.6b while using about a quarter of its active parameters. It combines a multilingual mmBERT encoder with ColBERT-style token matching, supports up to 8,192 tokens, and scores 63.33 on MNanoBEIR — between Qwen3-Embedding-8B (62.87) and Nemotron-3-Embed-8B (64.28).
Related event: LightOn's 115M-Parameter mLateOn Tops Multilingual Retrieval Benchmarks(4 posts)→
More from Research
- Five pharma giants federally train OpenFold3 on 20,000 private structures, +11 pts — jiqizhixin · 2026-10-16
- SWAN: a semantic-level text watermark that survives paraphrasing — burkov · 2026-10-16
- Interp researchers push back: 'networks are already clean' claim rests on dropout artifact — aryaman2020 · 2026-10-16
- LayerTrim: training-free pruning keeps reasoning LLMs intact on AIME — SeoulNatlUniv · 2026-10-16
- Simular presents neuro-symbolic computer use agents at BayLearn — xwang_lk · 2026-10-16
- Google's Hartmut Neven disputes chip test, outlines plan to link a brain to a quantum computer — Chris_Armstrong · 2026-10-16