Back to Gap Map Model components / Base / pretrained models

Snowflake Arctic

Snowflake
open weights / Overall score: 1.7

Snowflake Arctic is a 480B-parameter language model that pairs a 10B dense transformer with a residual mixture-of-experts layer of 128 experts, activating 17B parameters per token. Snowflake's AI research team trained it from scratch and released base and instruct checkpoints in April 2024, tuned for SQL generation, coding and instruction following.

Arctic has had no new release since April 2024. Snowflake's later Arctic-named models are embedding, text-to-SQL and speculative-decoding checkpoints, and Arctic Embed is its own entry.

Openness

3 high confidence
3.0
weights
open (base and instruct checkpoints, ungated on Hugging Face)
data
closed (the corpus is not released; the README points to a blog series on data rather than a dataset)
code
partial (inference and vLLM tutorials; the repository's training code targets Llama 3.1 fine-tuning, not Arctic pretraining)
license
Apache-2.0 (OSI)

Arctic's base and instruct weights are published under Apache-2.0 without a gate. Snowflake describes its data approach in blog posts but releases neither the training corpus nor the pretraining code, so it stays at open weights.

Adoption

1 high confidence
1.0

Hugging Face downloads summed across the base and instruct checkpoints.

Capability

2 medium confidence
2.0

Snowflake reports Arctic matching Llama 3 70B on coding, SQL and instruction following, but on general understanding it claims only to stay competitive, which places it with the 2024 open models of the Llama 3 8B class where InternLM 2.5 also sits. It is well below current releases.

  • https://raw.githubusercontent.com/Snowflake-Labs/snowflake-arctic/main/README.md recorded 2026-11-12

    README: "Arctic is on par or better than both LLAMA 3 8B and LLAMA 2 70B on enterprise metrics, while using less than ½ of the training compute budget" and "despite using 17x less compute budget, Arctic is on par with Llama3 70B in enterprise metrics like Coding (HumanEval+ & MBPP+), SQL (Spider) and Instruction Following (IFEval)", "while remaining competitive on overall performance".

Verified 2026-11-12, except some scores, still awaiting re-confirmation