Snowflake Arctic

Arctic came out of Snowflake AI Research in April 2024 as an enterprise-focused open LLM with an unusual dense-MoE hybrid design: a 10B dense transformer combined with a residual 128-expert MoE layer at 3.66B parameters per expert using top-2 gating, totaling 480B parameters with only 17B active per token. The design targeted cheap training relative to model quality, and Snowflake pitched it at enterprise intelligence tasks: SQL generation, coding, and instruction following. Everything shipped under Apache-2.0 with ungated weights on Hugging Face, plus a series of research cookbooks covering mixture-of-experts modeling, training systems, and data recipes, and inference tutorials for Hugging Face Transformers and vLLM; later repository work added fine-tuning support for Llama 3.1 405B and companion ArcticTraining and ArcticInference tooling. Serving 480B parameters, even sparsely activated, demands multi-GPU datacenter hardware, so adoption skews toward research on MoE architectures rather than self-hosted production, and it remains a well-documented open artifact for studying efficient MoE training at scale.