Snowflake's latest Arctic text embed model outperforms the competition

Snowflake today announced the open-source release of its Arctic embed family of AI models. Snowflake’s largest Arctic embed model outperforms the competition on the Massive Text Embedding Benchmark (MTEB) despite being significantly smaller. The largest Arctic embed model with 330 million parameters posted a record retrieval performance of 55.9. This marks a breakthrough for enterprises seeking powerful yet cost-effective search solutions.

The Arctic embed models are available now on Hugging Face under the Apache 2.0 license, and will soon be accessible within Snowflake’s Cortex embed function (currently in private preview).

Key Highlights:

• Best-in-Class Retrieval: The suite of five Arctic embed models delivers state-of-the-art search results across a wide range of sizes, outperforming other open-source models on the MTEB retrieval benchmark.
• Performance and Efficiency: Snowflake’s large (l) model uniquely beats closed-source models with 4x the parameters, offering superior performance with a smaller footprint.
• Long Document Support: The medium (m) model includes a long-context version for seamless handling of extended documents (up to 8192 tokens).
• Optimized for Enterprise: Arctic embed models excel in lower latency and reduced total cost of ownership (TCO) compared to similar-quality alternatives.

The Search for Better Search

Snowflake’s AI efforts are bolstered by its acquisition of Neeva. These models leverage Snowflake’s established search expertise and cutting-edge research, offering an unparalleled solution for enterprises looking to revamp their search capabilities with proprietary data and LLMs.

Open Source, Enterprise Grade

Snowflake’s Arctic embed models are poised to disrupt the market, challenging the dominance of closed-source embedding APIs. They offer a powerful open-source alternative to providers like OpenAI and Cohere, enabling deployment without vendor lock-in.