Arctic-Extract has credible specialist capability: Snowflake documents a proprietary vision model behind AI_EXTRACT for structured extraction across documents, images, and text, with fine-tuning support. Snowflake reports 0.9433 on DocVQA plus SQuAD v2 results, supporting a technical score near Gemma 3 but not broad frontier-model levels. Its document-centric visual I/O is strong, though narrower than Gemini 3.5 Flash’s broader multimodal score. It is not a coding or general-purpose agent model. DeepSWE lists no result; no SWE-bench result was found; LiveCodeBench exposes none; and the technical report evaluates document QA, entities, and tables instead. That warrants materially lower coding scores than Gemma 3, Grok 3, Kimi K2.5, GPT-5.5, and Claude Opus 4.8. SQuAD and extraction evaluations provide some evidence for task-specific language understanding, not general reasoning parity with those models. Cortex integration, SQL AI_EXTRACT, and documented fine-tuning make the developer workflow comparatively accessible for Snowflake users. However, supplied sources provide no direct Arctic-Extract price, throughput, latency, context-window, independent preference, or public adoption measurements; Artificial Analysis tracks no such model. Cost, speed, adoption, and pricing scores therefore remain conservative. Evidence quality is moderate: official documentation is clear, but performance claims are provider-reported and external benchmark coverage is thin.
