Snowflake is launching its own LLM: Snowflake Arctic. A talented expert in multiple domains, this generative AI model is required to perform complex tasks tailored to enterprises and integrates seamlessly with Snowflake’s data platform.

Snowflake introduces Snowflake Arctic. The Arctic is her own Great language model (LLM), which will compete with GPT-4, Llama 3 8B and other major LLMs. To this end, Snowflake provides some interesting assets to its Arctic model.

First of all, Snowflake Arctic is designed for enterprises and is part of the Snowflake data platform. The LLM therefore has secure and controlled access to company data without companies having to worry that their data will accidentally reach third parties.

Several experts

In addition, Snowflake opts for a so-called Arctic Mixture of experts (MoE) approach. This implies that Arctic is not a homogeneous model, but rather a combination of optimized sub-models. Briefly compare the approach to your brain, where certain zones are better at certain tasks than others. Snowflake claims that through this targeted, enterprise-grade optimization, Arctic outperforms the competition on many key workloads. The model can therefore generate SQL code very well.

Integration with the Snowflake platform then ensures that customers can use Arctic in a relatively accessible manner. Finally, the model fits into the Snowflake Cortex solution, which Snowflake previously introduced to give customers quick access to AI models. Cortex essentially ensures that you can run Arctic serverless with your own data.

Snowflake said its experts were able to build Arctic in three months, which is about an eighth of the cost normally associated with training a model of this size. The company makes Arctic available under an open Apache 2.0 license. Anyone who wants can use implementation templates.

AI focus

Snowflake isn’t ready for its test piece yet. Last year, the data cloud specialist jumped on Nvidia’s AI bandwagon with great enthusiasm. This was a smart and necessary move: Snowflake wants to manage and access the world’s data in a secure way, so that organizations can not only gain insights from their own data in a structurally sound way, but also combine it with data from external parties in a data cloud. All of this data would be harmful to Snowflake in the current hype climate without AI capabilities.

Meanwhile, we saw AI acceleration with Nvidia, as well as the introduction of more focused LLMs like Arctic Embed: a range of smaller, specialist models. Snowflake remains committed to supporting third-party models.