snowflake-arctic-embed-l-v2.0

by Snowflake

Multilingual text embeddings with Matryoshka truncation for cheaper retrieval

1.03Mdl/month

568Mparams

HuggingFace Run on your data, free

Identifiers

Model ID

Snowflake/snowflake-arctic-embed-l-v2.0

Feature URI

mixpeek://text_extractor@v1/snowflake_arctic_embed_l_v2

Overview

Arctic Embed 2.0 (Large) is Snowflake's open retrieval embedding model. Its headline trait is that it is multilingual without giving up English quality — it stays competitive on English MTEB retrieval while adding strong cross-lingual retrieval (MIRACL, CLEF) in one model, so you don't run a separate encoder per language.

Its second trait is practical: it is trained with Matryoshka Representation Learning, so a full 1024-dim vector can be truncated to 256 dims with only a few points of retrieval loss. That is a 4× cut in vector storage and memory with no re-embedding — you store the short vector and search it directly. On Mixpeek, Arctic Embed 2.0 is a strong default for multilingual document and transcript search where storage budget matters.

Architecture

~568M-parameter multilingual encoder (GTE-multilingual base) producing 1024-dim embeddings with an 8192-token context window. Trained with Matryoshka Representation Learning so the first 256 / 512 dimensions are themselves usable embeddings, plus contrastive training with hard negatives and a multilingual + English data mix tuned to avoid the usual multilingual-vs-English quality tradeoff. Mean pooling over token states; compare with cosine.

Mixpeek SDK Integration

import { Mixpeek } from "mixpeek";

const mx = new Mixpeek({ apiKey: "API_KEY" });

// Managed: create a collection over a bucket; Mixpeek runs this model's extractor
const collection = await mx.collections.create({
  namespace_id: "my-namespace",
  collection_name: "my-collection",
  source: { type: "bucket", bucket_ids: ["bkt_your_bucket"] },
  feature_extractor: {
    feature_extractor_name: "text_embedding",
    version: "v1",
    parameters: { model_id: "Snowflake/snowflake-arctic-embed-l-v2.0" },
  },
});

Capabilities

Strong English retrieval AND broad multilingual / cross-lingual retrieval in one model
Matryoshka dims: truncate 1024 → 256 for ~4× smaller vectors, minimal recall loss
8192-token context for long documents
Open weights (Apache-2.0), self-hostable
Drop-in 1024-dim dense vectors for any vector store

Use Cases on Mixpeek

Multilingual RAG: index documents in many languages and query in any of them with one encoder

Storage-constrained search: truncate to 256 dims to fit far more vectors per GB without re-embedding

Cross-lingual retrieval: an English query surfaces relevant non-English passages, and vice versa

Long-document transcript and report search within the 8192-token window

Benchmarks

Dataset	Metric	Score	Source
MTEB Retrieval (English)	nDCG@10	-	Competitive with strong English-only models of similar size; see report for the full table.
MIRACL (multilingual)	nDCG@10	-	Strong multilingual retrieval without an English-quality regression — the model's core claim.