Skip to main content
Back

Snowflake's Arctic Embed 2.0 Squeezes 74-Language Search Into 128 Bytes

Snowflake's open-source multilingual embedding model hits 874k downloads with strong retrieval across 74 languages, 8k context, and 128-byte compressed vectors.

  • Snowflake's Arctic Embed L v2.0 hits 874k downloads, 3,800 likes on Hugging Face.
  • 568M-parameter multilingual embedding model covering 74 languages under Apache 2.0.
  • Scores 55.6 BEIR, 55.8 MIRACL, 54.3 CLEF Full, beating bge-m3 on English tasks.
  • Matryoshka + 4-bit quantization compresses vectors to 128 bytes with under 3% quality loss.
  • 8,192-token context via RoPE, built on XLM-R Large backbone, drop-in for sentence-transformers.
  • Available via Ollama, vLLM Docker, ONNX, and Transformers.js.

Snowflake’s Arctic Embed L 2.0 has accumulated more than 874,000 downloads and 3,800 likes on Hugging Face at the snapshot reviewed for this article. The Apache 2.0 model has 568 million parameters, supports 74 languages, accepts up to 8,192 tokens, and produces 1,024-dimensional text embeddings. Its main appeal is practical: one encoder can handle English and multilingual retrieval while supporting aggressive vector compression.

Hugging Face download counters record file requests, including repeated and automated pulls, so they do not represent unique deployments. The numbers still show broad distribution for Snowflake’s open retrieval model.

One encoder, two deployment targets

Feedback on earlier Arctic Embed releases centered on multilingual support. Many teams have maintained separate English and multilingual encoders because multilingual models often lose accuracy on English retrieval. Arctic Embed 2.0 aims to consolidate those pipelines while preserving the English performance of the first-generation model.

Arctic Embed 2.0 variants
Variant Backbone Parameters Dimensions Context Design goal
arctic-embed-m-v2.0 GTE Multilingual 305M total, 113M excluding embeddings 768 8,192 tokens Lower inference cost
arctic-embed-l-v2.0 XLM-R Large derivative 568M total, 303M excluding embeddings 1,024 8,192 tokens Higher retrieval quality

Both variants use RoPE-based position handling for their extended context windows. The large model inherits from bge-m3-retromae and retains XLM-R-compatible interfaces, which broadens support across existing inference engines.

English holds steady as multilingual scores climb

Retrieval benchmarks commonly report nDCG@10, which rewards systems for placing relevant documents near the top of the first 10 results. The following Snowflake-reported scores are scaled to 100, with higher values indicating better rankings. BEIR covers 15 primarily English retrieval datasets, while MIRACL and CLEF evaluate multilingual retrieval.

Snowflake-reported average nDCG@10
Model Parameters Dimensions BEIR (15) MIRACL (4) CLEF Focused CLEF Full
arctic-l-v2.0 568M