Snowflake Inc. has launched a new large language model called Arctic, tailored for enterprise applications with capabilities in SQL and code generation.

Arctic uses a Dense-MoE Hybrid Transformer architecture, offering superior performance to other large language models while being more efficient in computing resource usage.

The company has made the model weights and training process public under an Apache 2.0 license, providing resources such as the 'Arctic Cookbook' for developers to create custom LLMs.

Snowflake's investment in Arctic amounted to three months of development, the use of 1,000 GPUs, and a $2 million budget, highlighting its commitment to competing in the AI space with rivals like Databricks.

Arctic is not only available for serverless inference through Snowflake Cortex but also on other platforms, with the Apache 2.0 license supporting integration into various AI-driven workflows.

Snowflake's CEO has positioned the release of Arctic as a step to advance open-source AI, responding to the growing enterprise need for customizable and efficient large language models.