Snowflake has added the open-source-based 'Arctic' to its data cloud platform. Snowflake's customers will be able to leverage data through an LLM that provides industry-leading openness.

Snowflake launched 'Snowflake Arctic,' an enterprise-grade Large Language Model (LLM) with industry-leading openness and performance, on the 24th.

Snowflake introduced Arctic, stating, "Designed with our proprietary Mixture-of-Experts (MoE) approach, Arctic supports best-in-class performance and productivity. Optimized for handling complex enterprise requirements, it meets top standards across various criteria, including SQL code generation and command execution."

Arctic established a new open standard for enterprise-grade AI technology under the Apache 2.0 license, which is free and commercially available, and Snowflake disclosed details of research on AI learning methods. In addition, Arctic LLM is part of the Snowflake Arctic model family and includes a text embedding model for search use.

Arctic also provides the flexibility to select inference and learning methods along with code templates. Users can use and customize Arctic with enterprise-preferred frameworks such as NVIDIA NIM, NVIDIA TensorRT-LLM, vLLM, and Hugging Face.

With Snowflake Cortex, you can also use Arctic immediately through serverless inference. Snowflake Cortex is a fully managed service that provides machine learning and AI solutions in the data cloud, along with a diverse list of models including Hugging Face, Lamini, Microsoft Azure, NVIDIA API Catalog, Perplexity, and Together AI. Arctic is also scheduled to become available on Amazon Web Services (AWS).

The Snowflake AI research team built Arctic in less than three months. They explained that Amazon Elastic Compute Cloud (Amazon EC2) P5 instances were used for model training, and the training cost was only one-eighth of that of similar models.

In addition, Arctic's differentiated MoE design can improve both training system and model performance through finely tuned data assembly tailored to enterprise requirements. You can activate 17 of the 480 billion parameters at a time.

Snowflake stated, "Arctic enables approximately 50% fewer parameters than DBRX and about 75% fewer parameters than Llama 3 70B. It surpasses the performance of representative existing open source models such as DBRX and Mixtral-8x7B in terms of coding (HumanEval+, MBPP+) and SQL generation (Spider), while demonstrating top-tier performance in general language understanding (MMLU, massive multi-tasking language understanding)."

Sridhar Ramaswamy, CEO of Snowflake, stated, “The Snowflake AI research team has created a significant turning point for the company by leading innovation at the forefront of the AI field. Snowflake is expanding the scope of possibilities for open-source AI by releasing industry-leading performance and efficiency to the AI community through open source. Snowflake’s AI capabilities are also growing, enabling us to provide customers with competent and trustworthy AI models.”