Research: A free 300M Gemma model predicts Gemini embeddings at 0.83 cosine similarity

dejanseo · x · 2026-08-14

Research demonstrates that a single 768×768 matrix can map vectors from Google's free, local Gemma-300m model onto the output of the paid gemini-embedding-001 API. On 50,000 held-out Wikidata entities, the mapped vector achieves a 0.83 cosine similarity to the true Gemini vector. A pure rotation recovers 0.77, while adding stretch and shear raises it to 0.831, effectively recovering the entity vector structure without API costs.

Original post →

More from Research

Research channel →