Snowflake Arctic

Arctic generates SQL queries from natural language instructions using a 10B dense transformer backbone combined with 128 expert MLP layers that selectively activate 17B parameters per token. The sparse MoE architecture routes SQL-generation tasks through specialized expert pathways trained on enterprise data patterns, enabling structurally-correct query generation for data warehouse operations. This is a primary optimization target, not a secondary capability.

Arctic generates and completes code across multiple programming languages by leveraging its 10B dense core and 128 expert MLP layers, with selective activation of 17B parameters per token. The mixture-of-experts routing mechanism directs code-generation tasks through specialized expert pathways trained on enterprise codebases and patterns, enabling context-aware code synthesis. Unlike general-purpose models, Arctic's training emphasizes enterprise code patterns and integration scenarios.

Arctic is released under Apache 2.0 license with ungated access to model weights and code. This permissive license allows unrestricted commercial use, modification, and redistribution without approval processes or usage restrictions. Developers can download weights directly, integrate into commercial products, and modify the model without licensing fees or vendor approval.

Arctic follows complex instructions and performs multi-step reasoning tasks by routing requests through its hybrid dense-MoE architecture, where the 10B dense backbone provides foundational instruction understanding and 128 expert layers specialize in enterprise-specific instruction patterns. The model activates 17B parameters per token, allowing selective expert engagement for different instruction types. Training emphasizes enterprise intelligence tasks (SQL, code, data analysis) while maintaining general instruction-following capability.

Arctic implements sparse mixture-of-experts inference through selective activation of expert pathways, where only 17B of 480B total parameters are active per token. The architecture combines a 10B dense transformer backbone with 128 expert MLP layers, using a gating mechanism to route tokens to relevant experts based on task characteristics. This sparse activation reduces computational cost and latency compared to dense models while maintaining performance through expert specialization.

Arctic is natively integrated into Snowflake Cortex, enabling inference directly within Snowflake's data cloud without data movement or external API calls. Queries can invoke Arctic through Cortex functions, allowing SQL-based access to the model for text generation, SQL generation, and code generation tasks. This integration eliminates data exfiltration concerns and enables seamless combination of model outputs with warehouse data operations.

Arctic is available as Apache 2.0 licensed open weights across multiple deployment platforms including Hugging Face, AWS, Azure, NVIDIA API Catalog, Replicate, Together, and Snowflake Cortex. The same model weights and code are used across all platforms, enabling consistent behavior and performance regardless of deployment choice. Developers can download weights directly or access via managed APIs, with inference frameworks like vLLM and TRT-LLM supported.

Arctic supports parameter-efficient fine-tuning using LoRA (Low-Rank Adaptation), allowing adaptation to domain-specific tasks without full model retraining. LoRA adds trainable low-rank matrices to frozen model weights, reducing memory and compute requirements for fine-tuning. Snowflake provides 'Training and Inference Cookbooks' documenting LoRA fine-tuning approaches, and offers a 'Build custom models with AI experts' service for business-specific customization.

Snowflake provides a complementary Arctic Embedding model optimized for semantic search and RAG (Retrieval-Augmented Generation) tasks. Arctic Embedding generates dense vector representations of text, enabling semantic similarity search, document retrieval, and RAG pipelines. The embedding model is designed to work in conjunction with Arctic for end-to-end AI workflows combining retrieval and generation.

Snowflake developed Arctic with an 'open data recipe' approach, achieving training cost under $2M while reaching competitive performance. The open data recipe is made available to users, enabling transparent understanding of training methodology and enabling reproduction or adaptation of the training process. This approach democratizes large-scale model training by documenting efficient training practices.

Arctic is positioned as a 'leader in enterprise tasks' based on benchmarks combining SQL generation, code generation, and instruction-following performance. The model is evaluated against alternatives like DBRX, Llama 3 70B, Mixtral 8x22B, and Mixtral 8x7B on enterprise-specific tasks. Snowflake provides benchmark results demonstrating Arctic's superiority on enterprise workloads while maintaining general-purpose capability.