Snowflake launches Arctic, an open enterprise LLM to take on DBRX

Snowflake just open-sourced Arctic LLM, a massive enterprise-grade language model released under the Apache 2.0 license for both commercial use and research. Instead of trying to write poetry or handle general creative writing, Arctic is built specifically for corporate database tasks, automated SQL generation, and internal copilots.

Under the hood and performance

Arctic uses a Mixture of Experts (MoE) architecture with a total of 480 billion parameters. However, it only activates 17 billion parameters at any given time to query its 128 specialized expert networks.

This sparse setup made training surprisingly fast and cheap. Snowflake trained Arctic in about three months using roughly 1,000 GPUs for around $2 million, pulling from public web datasets like RefinedWeb, C4, RedPajama, and StarCoder. The company claims training costs were roughly one-eighth of comparable models.

In terms of raw output, Snowflake claims Arctic beats Databricks’ DBRX in enterprise coding and SQL query generation. It also reportedly tops Meta’s Llama 2 70B and Mistral’s Mixtral-8x7B in those categories, though it doesn't beat Meta's newer Llama 3 70B. It also posts competitive scores on the standard MMLU reasoning benchmark.

Hardware requirements and limits

While Arctic is fully open source, it is far from lightweight. Fine-tuning or self-hosting the model locally requires around eight enterprise GPUs, putting local deployment out of reach for average developers.

Its context window is also relatively modest. Depending on how it is fine-tuned, Arctic handles between 8,000 and 24,000 words. That is noticeably smaller than giants like Gemini 1.5 Pro or Claude 3 Opus, meaning it can lose context across long conversational threads and still suffers from statistical hallucinations.

For those who want to run it, Snowflake released the model weights and coding templates on Hugging Face, Microsoft Azure, Lamini, and Together AI, alongside first-party hosting on Snowflake Cortex.

Why it matters

If your team spends half its day debugging complex database queries or building text-to-SQL interfaces, Arctic gives you a commercially viable, open-weight alternative to proprietary APIs like GPT-4.

That said, the steep hardware requirement means Arctic functions largely as an ecosystem funnel for Snowflake Cortex and supported cloud partners. Unless you have the infrastructure and compliance mandates to justify self-hosting an MoE giant, running it via managed cloud hosting will be the practical route for most businesses.