Lesson 2: Choosing an embedding model: leaderboards, prefixes, limits
28 min read·13 Oct 2026
Module 4 · Lesson 2 · About 50 minutes · Needs: Lesson 1 of this module (bge-small, cosine scores, normalising)
The situation
Kwame has opened the MTEB leaderboard. bge-small, the model behind Hollin's prototype, sits 84th. The top model scores 70.0 on retrieval against bge's 53.9, and he wants to switch. Owen asks whether a hosted embedding API would be cheaper than running a model ourselves. And a customer has written to the help centre in Polish.
Grace wants one answer she can sign off: which model, and how do we know? The running question is q32, "can I descale my C70 with vinegar". Its answer sits in the last section of the C70 user guide, the longest document Hollin has. The Hollin data is simulated.