Snowflake Arctic Instruct

Snowflake Arctic Instruct is a large language model developed by the Snowflake AI Research Team and released in April 2024. It combines a 10B dense transformer with residual 128×3.66B MoE MLP layers, resulting in 480B total parameters with only 17B active parameters selected via top-2 gating.

Arctic specializes in enterprise tasks — SQL generation, code generation, and instruction following — achieving top benchmark scores while requiring a fraction of the compute of comparable dense models. The architecture enables performant inference through optimal expert utilization.

Arctic is released under Apache 2.0 with ungated model weights. It is available on Snowflake Cortex, Hugging Face, AWS, Microsoft Azure, NVIDIA API Catalog, Replicate, and Together AI. Recommended inference hardware is an 8×H100 cluster (e.g. AWS p5.48xlarge). Fine-tuning is supported via LoRA and other PEFT methods.