New Open-Source SOTA for French Sparse Retrieval Beats BM25 on Synonyms
antoine_chaffin · x · 2026-10-12
A new open-source SOTA for French sparse retrieval was released. Sparse embedding models can retrieve synonyms for query words, unlike exact-match BM25. LightOn researcher Antoine Chaffin celebrated that their open models and data (mLateOn) helped create it: "The future is sparse" — and open releases call for open release.
More from Models
- DeepSeek desktop agent app reviewed: model promising, harness trails Cursor in 5 ways — casper_hansen_ · 2026-10-13
- Arena raises $200M Series B at $3.1B valuation, launches Alignment Index — arena · 2026-10-13
- SubQ Snap launches: 384K-context decision model at $0.05 per million tokens — ryanallred · 2026-10-13
- Dev Calls Zhipu's GLM 5.3 Flash the Best Practical Open-Source Model — EAccelerate_42 · 2026-10-13
- Reflection unveils Beam, a 501B open-weight MoE trained with RL on 10.5K GB300 GPUs — thione · 2026-10-12
- Mistral Previews Large 4: Trillion-Parameter Model With Open Weights Planned for Late October — thione · 2026-10-12