Snowflake Arctic Embed

Snowflake Arctic Embed is a family of five open-source text embedding models released by Snowflake in April 2024, built on search expertise the company acquired with Neeva. Ranging from x-small to large, the family targets retrieval-augmented generation and semantic search. The 334M-parameter large model posted 55.9 average retrieval performance on the MTEB Retrieval leaderboard — the only sub-billion-parameter model to do so at the time, beating closed-source models roughly four times its size. A long-context medium variant supports 8192 tokens, and the later v1.5 and 2.0 releases added Matryoshka compression and multilingual coverage.

The honest picture: benchmark leadership claims are vendor-reported on MTEB retrieval specifically, and the embedding landscape has since moved on with newer architectures. The models are English-centric in v1 (multilingual arrived with 2.0). But everything ships under Apache 2.0 on Hugging Face, the small sizes run cheaply on modest GPUs, and the family remains a pragmatic default for self-hosted RAG retrieval.