Today, Snowflake unveiled Arctic, a large language model (LLM) crafted for complex enterprise tasks like SQL generation, code creation, and instruction adherence.
Marketed as the “most open enterprise-grade LLM,” Arctic employs a mixture of expert (MoE) architecture, efficiently achieving top benchmarks for enterprise workloads. It showcases competitive performance against standard models from Databricks, Meta, and Mistral in areas such as world knowledge, common sense, reasoning, and mathematical capabilities.
“This is a watershed moment for Snowflake, with our AI research team at the forefront of innovation,” said CEO Sridhar Ramaswamy. “By providing industry-leading intelligence and efficiency in an open manner, we’re expanding the potential of open-source AI. Our research with Arctic will significantly enhance our ability to deliver reliable and efficient AI to our customers.”
The launch serves as Snowflake's strategic move to compete with Databricks, which has been aggressive in its AI initiatives. Snowflake’s focus on AI has recently intensified after its acquisition of Neeva and Ramaswamy's appointment as CEO.
Arctic: Designed for Enterprise Workloads
As modern enterprises embrace generative AI, there’s a surge in developing applications like retrieval-augmented generation (RAG) chatbots, data copilots, and code assistants. While numerous models exist, few specifically target enterprise requirements—this is where Snowflake Arctic excels.
“We believe AI will enhance the development of end-to-end AI products. Our vision is to create an API that allows business users to directly interact with data, democratizing it across the enterprise. Arctic is a vital step towards realizing this vision,” said Ramaswamy in a recent briefing.
Arctic utilizes a Dense MoE hybrid architecture, segmenting parameters into 128 specialized expert subgroups. These experts process only the input tokens they are best equipped to handle, activating only 17 billion of the 480 billion parameters when responding to a query. This targeted approach ensures high performance with minimal computing power.
Benchmarks indicate that Arctic effectively tackles enterprise tasks, achieving an average score of 65% across various tests. This performance closely aligns with Llama 3 70B’s average enterprise score of 70% and trails only behind Mixtral 8X22B’s 70%.
In the Spider benchmark for SQL generation, Arctic scored 79%, surpassing Databricks' DBRX and Mixtral 8X7B, and coming close to Llama 3 70B and Mixtral 8X22B. For coding tasks, Arctic achieved 64.3%, outperforming Databricks and the smaller Mixtral model while falling short of Llama 3 70B and Mixtral 8X22B.
Notably, in the IFEval benchmark for instruction-following capabilities, Arctic scored 52.4%, outperforming most competitors except for the latest Mixtral model.
Efficiency and Cost-Effectiveness
Snowflake asserts that Arctic's level of enterprise intelligence was achieved with groundbreaking efficiency, utilizing a training compute budget of under $2 million—significantly less than other models like Llama 3 70B, which used 17 times more computing resources. Additionally, Arctic's use of only 17 active parameters further enhances its cost-effectiveness.
Availability Under the Apache 2.0 License
Snowflake is making Arctic accessible through Cortex, its LLM application development service, and across various model catalogs, including Hugging Face, Lamini, Microsoft Azure, Nvidia API catalog, Perplexity, and Together. Users can download Arctic's model weights and code from Hugging Face under an Apache 2.0 license, facilitating unrestricted personal, commercial, or research use.
In conjunction with the model release, Snowflake is providing a data recipe for efficient fine-tuning on a single GPU and comprehensive research cookbooks detailing the model's design and training processes.
“The cookbook is crafted to accelerate the learning curve for anyone interested in world-class MoE models, offering both high-level insights and detailed technical specifications to empower users to construct efficient and cost-effective LLMs like Arctic,” stated Baris Gultekin, Snowflake’s head of AI, during the press briefing.
Ranking
