MTEB@mteb.orgThe best embedding score available at the end of 2023 needed a 304M-parameter model. Today the same score comes from an 11M one — 28x fewer active parameters.5/8/26, 9:43 am·2 reposts·15 likes
regularfry.bsky.social@regularfry.bsky.social5 AugThe ~100M beating 2024's ~10B is no slouch either, if I'm reading the chart right1