Snowflake Arctic targets enterprise SQL and code generation
Snowflake's open-source Arctic model uses a Dense-Mixture-of-Experts architecture with 480B total parameters but only 17B active during inference, targeting enterprise SQL and code generation workloads under an Apache 2.0 license.
Score breakdown
Engineers building AI-powered database or coding tools have a domain-specialized, commercially permissive open-source alternative to general-purpose models, with deployment paths through Hugging Face, NVIDIA NIM, and Amazon SageMaker JumpStart.
- 01Arctic uses a Dense-Mixture-of-Experts (MoE) hybrid transformer architecture with 480B total parameters.
- 02Only 17B parameters are activated during inference via a top-2 gating mechanism.
- 03The architecture includes a 10B dense transformer combined with 128 specialized experts.
Snowflake's Arctic is an open-source LLM designed from the ground up for enterprise use cases, particularly SQL and code generation. Its Dense-Mixture-of-Experts (MoE) hybrid transformer architecture combines a 10B dense transformer with 128 specialized experts, yielding 480B total parameters. A top-2 gating mechanism ensures only 17B parameters are active at inference time, aiming to deliver the knowledge capacity of a very large model at the computational cost of a much smaller one — a meaningful efficiency advantage for teams deploying AI at scale.
This specialization is reflected in benchmark results: Arctic performs strongly on Spider for SQL generation, HumanEval+ and MBPP+ for code generation, and IFEval for instruction following.
The model's training curriculum was deliberately structured in three stages, with the latter two phases heavily weighted toward enterprise-focused data covering code, SQL, and STEM. This specialization is reflected in benchmark results: Arctic performs strongly on Spider for SQL generation, HumanEval+ and MBPP+ for code generation, and IFEval for instruction following. The instruct-tuned version is available on Hugging Face under the model ID `Snowflake/snowflake-arctic-instruct` and can be loaded via the `transformers` pipeline with `trust_remote_code=True`.
Beyond self-hosting, Arctic is available through NVIDIA NIM and can be deployed from Amazon SageMaker JumpStart. Snowflake released the model weights and code under an Apache 2.0 license with ungated commercial access, giving builders full transparency and the ability to customize the model for their specific needs.
Key facts
- 01Arctic uses a Dense-Mixture-of-Experts (MoE) hybrid transformer architecture with 480B total parameters.
- 02Only 17B parameters are activated during inference via a top-2 gating mechanism.
- 03The architecture includes a 10B dense transformer combined with 128 specialized experts.
- 04Training used a three-stage curriculum, with the latter two stages focused on code, SQL, and STEM data.
- 05Arctic benchmarks on Spider (SQL), HumanEval+, MBPP+ (code), and IFEval (instruction following).
- 06The instruct-tuned model is available on Hugging Face as `Snowflake/snowflake-arctic-instruct`.
- 07Released under Apache 2.0 with ungated access to weights and code; also available via NVIDIA NIM and Amazon SageMaker JumpStart.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 8, 2026 · 15:36 UTC. How this works →