Snowflake Inc. announced Snowflake Arctic, a large language model (LLM) uniquely designed to be the most open, enterprise-grade LLM on the market. With its unique Mixture-of-Experts (MoE) architecture, Arctic delivers top-tier intelligence with unparalleled efficiency at scale. It is optimized for complex enterprise workloads, topping several industry benchmarks across SQL code generation, instruction following, and more.

In addition, Snowflake is releasing Arctic?s weights under an Apache 2.0 license and details of the research leading to how it was trained, setting a new openness standard for enterprise AI technology. The Snowflake Arctic LLM is a part of the Snowflake Arctic model family, a family of models built by Snowflake that also include the best practical text-embedding models for retrieval use cases. Snowflake's AI research team, which includes a unique composition of researchers and system engineers, took less than three months and spent roughly one-eighth of the training cost of similar models when building Arctic.

Trained using Amazon Elastic Compute Cloud (Amazon EC2) P5 instances, Snowflake is setting a new baseline for how fast open, enterprise-grade models can be trained, ultimately enabling users to create cost-efficient custom models at scale. As a part of this strategic effort, Arctic?s differentiated MoE design improves both training systems and model performance, with a meticulously designed data composition focused on enterprise needs. Arctic also delivers high-quality results, activating 17 out of 480 billion parameters at a time to achieve quality with unprecedented token efficiency.

In an efficiency breakthrough, Arctic activates roughly 50% less parameters than DBRX, and 75% less than Llama 3 70B during inference or training. In addition, it outperforms leading open models including DBRX, Mixtral-8x7B, and more in coding (HumanEval+, MBPP+) and SQL generation (Spider), while simultaneously providing leading performance in general language understanding (MMLU).