Arctic Instruct

Arctic Instruct is Snowflake's open-weights MoE model released April 2024, with 480B total parameters but only around 17B active per token via a hybrid dense-plus-sparse architecture — a 10B dense transformer backbone feeding 128 experts of roughly 3.66B each. It was designed specifically for enterprise SQL generation and retrieval-augmented generation tasks, where early internal benchmarks showed strong performance on structured data queries relative to similarly priced alternatives. The 4K context window is the most significant practical limitation and rules it out for most long-document workflows. Licensed Apache 2.0 with no commercial restrictions. Hosted coverage is sparse as of 2026; for most teams it is primarily a research reference or fine-tuning base rather than a production inference choice.