LightOn open-sources mSlideOn, best-ever sub-500M model on MTEB retrieval
tomaarsen · x · 2026-10-22
LightOn released the embedding model mSlideOn on Hugging Face (Apache 2.0, works with sentence-transformers). Per @tomaarsen, it is the best and second-best sub-500M model ever on MTEB (English, v2) Retrieval, matching the performance of its larger sibling.
The model is multilingual (English, French, German, Spanish, etc., plus code search), supports sparse retrieval and Matryoshka embeddings, is fine-tuned from mLateOn, and ships with filtered fine-tuning datasets for each language.
More from Models
- unsloth ships GGUF quantizations of RecursiveSI/flint on Hugging Face — unsloth · 2026-10-22
- Generalist AI's GEN-1.5 puck-packing bot shows emergent physical reasoning and backtracking — chris_j_paxton · 2026-10-22
- User suspects OpenAI's "Approve for me" feature is behind recent token-draining spike — ssh4net · 2026-10-22
- Dev names Kimi K2 his favorite LLM: won't sycophantically agree, may outcode Sonnet 4 — gnukeith · 2026-10-22
- 31,000-question study finds LLMs across 24 languages are secretly obsessed with Japan — dioscuri · 2026-10-22
- Wiz says Gemini 4 Argon helped find decade-old critical vulnerability, coordinated fix across industry — HaydnBelfield · 2026-10-22