Robots Atlas>ROBOTS ATLAS
Snowflake Arctic Embed M v1.5
AI Models›Snowflake Arctic

Snowflake Arctic Embed M v1.5

1.5 · Family: Snowflake Arctic
Snowflake Arctic Embed M v1.5 is a 109M-parameter embedding model optimized for vector compression. Supports MRL down to 128 bytes per vector retaining 98% quality. Apache 2.0 licensed.
✓ Active✓ Public access⚖ Open weightsEmbedding model📁 Snowflake Arctic
Context window
512
tokens
Parameters
109M
parameters
Release date
18 July 2024
🏢Snowflake Inc.Producer
Access:APIDownloadDeployment:☁ Cloud💻 Local

Overview

Snowflake Arctic Embed M v1.5 is an updated embedding model released in July 2024, featuring 109M parameters (BERT architecture) with 768-dimensional vectors and a 512-token context window. It achieves an MTEB Retrieval score of 55.14 NDCG@10.

Its defining feature is vector compressibility: via Matryoshka Representation Learning (MRL), vectors can be truncated to 256 dimensions (54.2 NDCG@10, retaining 99% quality), and combined with 4-bit scalar quantization, compressed to 128 bytes per vector (53.7 NDCG@10, retaining 98% quality), enabling 7.8M vectors per GB.

Available on Snowflake Cortex AI as `snowflake-arctic-embed-m` and `snowflake-arctic-embed-m-v1.5`, on Hugging Face, via Sentence Transformers, Transformers.js, and Text Embeddings Inference. Apache 2.0 licensed for free commercial use.

Classification
Embedding model
Family: Snowflake Arctic
Access & deployment
APIDownload
CloudLocal
Weights: Open weights
Key parameters
📏 Context: 512
🧩 Parameters: 109M
✓ Fine-tuning
📥 Input: text
Platforms
Snowflake Cortex AI

Technical specification

Context window
512
tokens
Parameters
109M
parameters
License
Apache 2.0
Features:✓ Fine-tuning
Modalities
⬇ Input
text
⬆ Output
structured_data

Capabilities and applications

Native model capabilities
Multilingual
Understanding and generating text in many languages.
Category: language

Benchmark results

2 benchmarks
MTEB Retrieval (BEIR-15)
NDCG@10 · No compression (768 dimensions, float32)
55.14points
📄 Hugging Face model card (self-reported)
MTEB Retrieval (BEIR-15) – 256-dim MRL
NDCG@10 · After MRL truncation to 256 dimensions
54.2points
📄 Hugging Face model card (self-reported)

Pricing

Deployment and security

☁ Available on platforms
☁Snowflake Cortex AIPlatform
🔒 Security / Enterprise
✓ Verified enterprise information

Models hosted in Snowflake Cortex AI use Snowflake's full security infrastructure. Data does not leave the platform's security boundary.

Updated: 28 Apr 2026↗ Security documentation

Sources and related pages