The data analytics company Databricks is on a mission to provide data intelligence to every enterprise by enabling organizations to understand and utilize their unique data to build their AI systems. Central to this mission is the ability to leverage a large language model tailored to meet the specific needs of enterprises. Recently, Databricks announced the release of DBRX, a new open, general-purpose large language model that aims to set new benchmarks for performance and efficiency in the industry.
DBRX was developed by Databricks’ Mosaic Research team, which the company acquired in June 2023 as part of its MosaicML acquisition. Built as a Mixture-of-Experts model, DBRX boasts impressive processing speed efficiency while maintaining high model performance. With 36 billion parameters out of 132 billion in operation at any given time, DBRX strikes a balance between speed and quality that outperforms existing open-source models and even surpasses OpenAI’s GPT-3.5 across various standard benchmarks.
One of the key features of DBRX is its customizability, allowing enterprises to tailor the model to their unique data and requirements. The open-source license further enables organizations to adjust DBRX to their specific needs, potentially leading to improved performance compared to proprietary models. DBRX was entirely developed on the Databricks platform, utilizing tools like Unity Catalog, Apache Spark, and Mosaic AI Training for data governance, processing, and model training, respectively, to ensure a seamless integration.
Databricks ensures that DBRX is accessible to its customers through APIs, enabling easy integration into existing workflows and applications within the Databricks environment. Customers also have the flexibility to pre-train their DBRX-class models from scratch or continue training on top of Databricks-provided checkpoints, allowing them to adapt the model’s capabilities to their enterprise-specific needs. Early rollouts of DBRX in Databricks’ GenAI-powered products have shown promising results, especially in applications like SQL query generation and optimization.
The release of DBRX aligns with the trend of open large language models tailored for enterprise needs, following similar initiatives from other major players in the market. The MoE architecture used in DBRX addresses the challenge of resource-intensive operations in AI model deployment, making it a powerful and economically viable option for a broader range of enterprises. By outperforming established models like GPT-3.5 on various benchmarks, Databricks solidifies its position as a significant player in the AI space.
As organizations look to incorporate advanced AI capabilities into their operations, DBRX offers a direct pathway to enhance tasks automation, data analysis, and decision-making processes. Databricks’ commitment to democratizing AI through the release of DBRX underscores the company’s dedication to driving AI-enabled transformation across industries. By open-sourcing DBRX, Databricks opens opportunities for the global AI research community to contribute to and improve the model, fostering a collaborative ecosystem around its AI technologies.
