Tiny 307M-parameter model outperforms 26x larger Qwen in embedding benchmarks

lateinteraction · x · 2026-08-27

An experiment demonstrates that a tiny 307M-parameter mLateOn model, using Late Interaction, beats every single-vector method including the 26x larger Qwen3-Embedding-8B in zero-shot nDCG. The finetuned mLateOn-med can handle hundreds of millions of tokens with an index size smaller than Qwen3's, showcasing Late Interaction's efficiency.

Related event: Tiny 307M Late Interaction Model Beats Embedders 26x Its Size(3 posts)→

Original post →

More from Models

Models channel →