Arctic Embed
Snowflake'/%3e%3cdefs%3e%3cpattern%20id='pattern0_1240_293949'%20patternUnits='userSpaceOnUse'%20patternTransform='matrix(11.2%200%200%2011.2%203.93617%200)'%20preserveAspectRatio='none'%20viewBox='0%200%2016%2016'%20width='1'%20height='1'%3e%3cg%20id='pattern0_1240_293949_inner'%3e%3crect%20width='16'%20height='16'%20fill='%23f88376'%20mask='url(%23hatch_gaps)'/%3e%3cmask%20id='hatch_gaps'%20maskUnits='userSpaceOnUse'%20x='0'%20y='0'%20width='16'%20height='16'%3e%3crect%20width='16'%20height='16'%20fill='white'/%3e%3crect%20x='6'%20width='2'%20height='2'%20fill='black'/%3e%3crect%20x='6'%20y='8'%20width='2'%20height='2'%20fill='black'/%3e%3crect%20x='4'%20y='2'%20width='2'%20height='2'%20fill='black'/%3e%3crect%20x='4'%20y='10'%20width='2'%20height='2'%20fill='black'/%3e%3crect%20x='2'%20y='4'%20width='2'%20height='2'%20fill='black'/%3e%3crect%20x='2'%20y='12'%20width='2'%20height='2'%20fill='black'/%3e%3crect%20y='6'%20width='2'%20height='2'%20fill='black'/%3e%3crect%20y='14'%20width='2'%20height='2'%20fill='black'/%3e%3crect%20x='14'%20width='2'%20height='2'%20fill='black'/%3e%3crect%20x='14'%20y='8'%20width='2'%20height='2'%20fill='black'/%3e%3crect%20x='12'%20y='2'%20width='2'%20height='2'%20fill='black'/%3e%3crect%20x='12'%20y='10'%20width='2'%20height='2'%20fill='black'/%3e%3crect%20x='10'%20y='4'%20width='2'%20height='2'%20fill='black'/%3e%3crect%20x='10'%20y='12'%20width='2'%20height='2'%20fill='black'/%3e%3crect%20x='8'%20y='6'%20width='2'%20height='2'%20fill='black'/%3e%3crect%20x='8'%20y='14'%20width='2'%20height='2'%20fill='black'/%3e%3c/mask%3e%3c/g%3e%3c/pattern%3e%3c/defs%3e%3c/svg%3e)
Arctic Embed is a family of retrieval embedding models sized from xs to l. The 2.0 releases build on the BGE-M3 RetroMAE backbone to handle multilingual retrieval at up to 8,192 tokens without giving up English quality. Matryoshka representation learning and quantization-aware training let a vector be truncated to 256 dimensions or compressed to 128 bytes per document. Snowflake develops and publishes the family.
The v1 xs checkpoint carries no license field on the Hub while every other family member and the family card declare Apache-2.0; the score records Apache-2.0 on the card's family-wide statement.
Openness
3 high confidence- weights
- open(ungated checkpoints at xs, s, m, m-long, m-v1.5, l, m-v2.0 and l-v2.0)
- data
- closed(described only as roughly 400M pairs mixing public datasets with proprietary web-search data, then about 1M mined triplets
- code
- partial(the arctic-embed repository ships inference and embedding-compression notebooks, not a training pipeline)
- license
- Apache-2.0(OSI)
Arctic Embed ships open weights under Apache-2.0, but about half its pretraining mixture is proprietary web-search data that Snowflake describes without releasing. The public repository offers only inference and compression notebooks rather than a training pipeline, which keeps it at open weights rather than open source.
- https://huggingface.co/api/models/Snowflake/snowflake-arctic-embed-l-v2.0 recorded 2026-09-11
license: apache-2.0; gated: false; private: false; downloadable checkpoint files present.
- https://huggingface.co/api/models/Snowflake/snowflake-arctic-embed-m-v2.0 recorded 2026-09-11
license: apache-2.0; gated: false; private: false.
- https://huggingface.co/api/models/Snowflake/snowflake-arctic-embed-s recorded 2026-09-11
license: apache-2.0; gated: false; private: false.
- https://huggingface.co/api/models/Snowflake/snowflake-arctic-embed-xs recorded 2026-09-11
gated: false; private: false; no license field and no license tag on the v1 xs checkpoint.
- https://huggingface.co/Snowflake/snowflake-arctic-embed-l-v2.0/raw/main/README.md recorded 2026-09-11
States "Arctic is licensed under the Apache-2. The released models can be used for commercial purposes free of charge."
- https://raw.githubusercontent.com/Snowflake-Labs/arctic-embed/main/README.md recorded 2026-09-11
Describes pretraining on about 400m samples of "a mix of public datasets and proprietary web search data" followed by about 1m mined triplets, and offers inference and quantization notebooks rather than training code.
Adoption
4 high confidenceAdoption is measured as Hugging Face downloads summed across the eight checkpoints Arctic Embed ships, led by the l-v2.0 model. The s and m checkpoints add further volume without changing the reading.
- https://huggingface.co/api/models/Snowflake/snowflake-arctic-embed-l-v2.0 recorded 2026-09-11
downloads: 821744.
- https://huggingface.co/api/models/Snowflake/snowflake-arctic-embed-m recorded 2026-09-11
downloads: 397003.
- https://huggingface.co/api/models/Snowflake/snowflake-arctic-embed-s recorded 2026-09-11
downloads: 590982.
Capability
4 high confidenceArctic Embed is multilingual and long-context, and its published table puts it ahead of bge-m3 and gte-multilingual-base on English BEIR and CLEF while level with them on MIRACL. It sits a step below Harrier-OSS because its numbers are BEIR and CLEF rather than MMTEB, the benchmark this category treats as primary, and no successor to the 2.0 release has shipped since.
- https://huggingface.co/Snowflake/snowflake-arctic-embed-l-v2.0/raw/main/README.md recorded 2026-09-11
Quality table gives snowflake-arctic-l-v2.0 BEIR (15) 55.6, MIRACL (4) 55.8, CLEF (Focused) 52.9 and CLEF (Full) 54.3 average NDCG@10, against 48.8 / 56.8 / 40.8 / 41.3 for bge-m3 and 51.1 / 52.3 / 47.7 / 53.1 for gte; truncation to 256 dimensions costs under 3%.
- https://raw.githubusercontent.com/Snowflake-Labs/arctic-embed/main/README.md recorded 2026-09-11
Records the 12/04/2024 release of the l-v2.0 and m-v2.0 multilingual models as the most recent additions to the family.
Verified 2026-09-11