Snowflake Arctic Embed 2.0
Compact multilingual retrieval model with compression-friendly training
About
Arctic Embed 2.0 is Snowflake's second-generation open embedding family, built on the BGE-M3 retroMAE backbone for drop-in compatibility. The large variant has around 568M parameters, an 8192-token context window, and is trained to be multilingual without sacrificing English retrieval quality. Its standout engineering trait is compression-friendliness: Matryoshka training plus quantization-aware training lets it hold retrieval quality at as little as 128 bytes per vector, which matters a lot for large indexes. It is Apache 2.0 licensed and genuinely open. The honest caveats: independent third-party evaluations are thinner than for BGE or E5, and the query-prefixing setup adds a small integration wrinkle.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
MTEB Retrieval NDCG@10 ~0.55 (Snowflake-reported)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 6/10
Scored 6/10 (metrics tier, metrics-v1). Evidence: Apache 2.0, ~568M params, 8192 context, MRL compression-friendliness corroborated in Ollama library docs and Snowflake's release materials.; Release dated December 2024 (m-v2.0 dated Dec 4, 2024; l-v2.0 in the same window) — exact day for the large variant not pinned, left null.; Independent third-party evaluations are thinner than for BGE/E5 — metrics tier..