Snowflake Introduces Arctic: An Open ‘Mixture-of-Experts’ LLM to Compete with DBRX and Llama 3

Snowflake’s new offering, Arctic, represents a significant advancement in large language models (LLMs), particularly tailored for complex enterprise tasks like SQL and code generation. Unlike other LLMs, Arctic leverages a unique mixture of expert (MoE) architecture to excel in enterprise benchmarks while maintaining efficiency. Additionally, Arctic delivers competitive performance across various standard benchmarks, nearly matching other open models from industry players like Databricks, Meta, and Mistral in tasks involving world knowledge, common sense, reasoning, and mathematical capabilities. This makes Arctic a promising option for organizations seeking advanced language capabilities for their enterprise workloads.

The launch of Arctic marks a significant milestone for Snowflake, showcasing their commitment to innovation in the AI space. CEO Sridhar Ramaswamy emphasized the importance of delivering industry-leading intelligence and efficiency to the AI community in an open manner. This move positions Snowflake to compete more effectively with industry players like Databricks, known for their aggressive AI efforts in the data platform space. Snowflake’s recent acceleration in AI endeavors, fueled by the acquisition of Neeva and Ramaswamy’s leadership, underscores their dedication to staying competitive and delivering reliable, efficient AI solutions to their customers.

Snowflake Arctic is poised to address the growing demand for generative AI applications in modern enterprises by offering a specialized solution tailored to enterprise tasks. With the rise of applications like retrieval-augmented generation (RAG) chatbots, data copilots, and code assistants, there’s a need for models optimized for the unique requirements of business workflows. By providing a platform optimized for complex enterprise workloads such as SQL generation, code generation, and instruction following, Snowflake Arctic aims to empower enterprises to harness the full potential of generative AI in their operations. This specialized focus on enterprise tasks distinguishes Snowflake Arctic from other models in the market, positioning it as a valuable tool for organizations looking to leverage AI to drive innovation and efficiency.

Snowflake Arctic’s Dense MoE hybrid architecture represents a significant advancement in AI technology, enabling more efficient and targeted performance for enterprise tasks. By dividing parameters into fine-grained expert subgroups and employing a dynamic data curriculum for training, Arctic ensures that only select parameters of the model are activated in response to a query. This approach minimizes compute consumption while delivering high-performance results.

According to benchmarks shared by Snowflake, Arctic achieves an average score of 65% across multiple tests, demonstrating its effectiveness in handling enterprise tasks. This places Arctic on par with other leading models like Llama 3 70B and slightly behind Mixtral 8X22B. With its focus on optimized performance and minimal compute consumption, Arctic is poised to play a crucial role in advancing Snowflake’s vision of democratized data within the enterprise, enabling business users to directly interact with data and drive innovation.

Snowflake Arctic demonstrates impressive performance across various benchmarks, showcasing its capabilities in SQL generation, coding tasks, and instruction following. In the Spider benchmark for SQL generation, Arctic outperformed competitors like Databricks’ DBRX and Mixtral 8X7B, and nearly matched leading models like Llama 3 70B and Mixtral 8X22B. Similarly, in coding tasks, Arctic surpassed Databricks and smaller Mixtral models, although it trailed behind Llama 3 70B and Mixtral 8X22B.Of particular note is Arctic’s performance in the IFEval benchmark, where it scored 52.4%, demonstrating strong instruction following capabilities compared to most competitors, except the latest Mixtral model.What sets Arctic apart is its breakthrough efficiency, achieved with a training compute budget of under $2 million, significantly less than other open models like Llama 3 70B. Arctic also utilizes only 17 active parameters to achieve these results, driving further cost benefits.

Snowflake is making Arctic available through its Cortex LLM app development service and various model gardens and catalogs, including Hugging Face, Lamini, Microsoft Azure, Nvidia API catalog, Perplexity, and Together. Arctic’s model weights and code can be downloaded under an Apache 2.0 license, enabling ungated use for personal, commercial, or research applications.In addition to releasing model weights and codes, Snowflake is providing a data recipe for efficient fine-tuning on a single GPU and comprehensive research cookbooks, offering insights into Arctic’s design and training process. This effort aims to expedite the learning process for building LLMs like Arctic, ensuring anyone can efficiently and economically develop their desired intelligence.⬤