mLateOn wins HAKARI-Bench by 8.33 points with a quarter of runner-up's active params
IgorCarron · x · 2026-08-20
HAKARI-Bench creator @hotchpotch reports that LightOn's mLateOn tops all 11 evaluated late-interaction retrievers: 65.52 Overall Macro, 8.33 points above runner-up pplx-embed-v1-late-0.6b (57.19), with only 115.1M active parameters versus 440.6M. With 312M total params, it scores 63.33 on MNanoBEIR — in the same tier as 8B dense models like Qwen3-Embedding-8B and Nemotron-3-Embed-8B. It supports 8,192-token inputs, reusable document encodings for reranking, and stays top-tier on English retrieval.
Related event: LightOn's 115M-Parameter mLateOn Tops Multilingual Retrieval Benchmarks(4 posts)→
More from Research
- UCSD Researchers Propose Dataset Watermarking Detectable via Black-Box API Access — yuxiangw_cs · 2026-10-16
- EmbodiedEvo tops RoboDojo as agents inherit reusable skills and lessons — Aiden_Tech_Ai · 2026-10-16
- CHI reviewer laments: papers' Discussion sections now cite nothing — IanArawjo · 2026-10-16
- Casper: Understanding 90% of neurons ≠ understanding the network — StephenLCasper · 2026-10-16
- NVIDIA Launches cuPotential in CUDA-X to Bring AI Agents to Semiconductor Materials Simulation — tomaszbednarz · 2026-10-16
- Agentic Search Systems Are Rewarded Only for Final Answers, Raising Faithfulness Concerns — burkov · 2026-10-16