Fordel StudiosWelcome EmbeddingGemma, Google’s new efficient embedding model
What Happened
Welcome EmbeddingGemma, Google’s new efficient embedding model
Fordel's Take
Google released EmbeddingGemma, a 7M-parameter encoder that cranks out 1k-dim English embeddings at 1.3 GB/s on a single T4.
Swapping it in for your RAG retriever cuts memory 4× versus text-embedding-ada-002 and keeps recall within 1 % on BEIR, so running Ada for vector lookup is just burning money.
Teams under 5k daily queries can ignore the swap; anyone above that should swap the endpoint this week and pocket the $200/month GPU saving.
What To Do
Swap to EmbeddingGemma endpoints instead of Ada-002 because the 7M model runs on a $0.05/hr T4 and keeps 99 % BEIR recall.
Cited By
React
Get the weekly AI digest
The stories that matter, with a builder's perspective. Every Thursday.