Free Lesson
Stop picking embedding models off the MTEB leaderboard
Part of AI Product Engineering
30 min
Jul 29, 2026 12:00 PM
Virtual (Zoom)
In this video
What you'll learn
Choose an embedder for your data and budget
Weigh the traits a leaderboard skips, including vector quantization, model size, and Matryoshka dimensions, so the model fits your latency and cost.
Why binary vectors work for most cases
See when binary vectors are enough, and how multi-stage ranking recovers the relevance that compression costs you.
Fine-tune the embedder on your own data
Use VespaTune, an open source UI, to fine-tune an embedder on your data instead of trusting a benchmark score.
Why this topic matters
It's tempting to pick an embedding model according to which one tops the MTEB leaderboard. But MTEB doesn't tell you how the model will do on your data, where you'll often need to fine-tune it, or how it performs on speed and quality once you change quantization and dimensions. Radu shows how to juggle these variables so you get the most for your money.
You'll learn from
Radu Gheorghe
Software Engineer, Vespa.ai
Hamel Husain
ML Engineer with 20+ years of experience