Elevating AI with Databricks on Azure: Introducing the Latest Large Language Models


In the ever-evolving landscape of artificial intelligence (), the integration of advanced models into accessible platforms is a game-changer for developers and businesses alike. Today we're excited to announce the availability of Databricks' new large language models, databricks/dbrx-base and databricks/dbrx-instruct, in the Microsoft Azure Model Catalog. This addition underscores our ongoing commitment to offering our customers a diverse selection of models. The new models are also available in Azure Databricks, on the Databricks marketplace, and as open models from HuggingFace.  

Databricks Meets Azure: A Synergy of Power and Flexibility 

Databricks is one of the companies at the forefront of the and analytics space, offering data intelligence to every enterprise by allowing organizations to understand and use their unique data to build their own AI systems. The introduction of their latest models, databricks/dbrx-base and databricks/dbrx-instruct, into the Microsoft Azure AI platform and Azure Databricks, is a testament to our joint commitment to advancing AI innovation across the machine learning lifecycle. 

Our partnership with Databricks is a collaboration aimed at enhancing cloud-based and AI, centering on the integration of Databricks's platform and tools with Microsoft Azure.  Azure Databricks, a Spark-based data analytics platform combined with Azure's purpose-built infrastructure provides a collaborative environment for data engineers, business analysts, and data scientists. Within this ecosystem, data managed in Azure Databricks is readily accessible via OneLake in Microsoft Fabric through intuitive shortcuts. This data can then be utilized in Azure AI Studio, enhancing the Large Language Models (LLMs) in the model catalog through the application of a Retrieval-Augmented Generation (RAG) pattern or by fine-tuning existing models, thus driving forward innovation and excellence in AI development.  

The availability of databricks/dbrx-base and databricks/dbrx-instruct within the Microsoft Azure AI model catalog opens a realm of possibilities.  Developers can now easily integrate these models into their projects, benefiting from Azure's scalability and Databricks' advanced data capabilities. This collaboration not only simplifies the development process but also encourages innovation, allowing for the creation of intelligent and intuitive applications. 

Exploring the New Models 

DBRX stands as a pioneering Mixture-of-Experts (MoE) model, crafted atop the innovations of the MegaBlocks research and open-source initiative from Databricks. This design propels the model to achieve remarkable speed in tokens per second, setting a precedent for future state-of-the-art open-source models to adopt MoE architectures. With a colossal 132B parameters, of which 36B are activated for any given input, DBRX demonstrates unparalleled capacity. The model underwent pre-training on a vast corpus of 12T tokens, sourced from meticulously selected data, with support for up to 32K tokens in context length. The construction of the training dataset leveraged the comprehensive array of Databricks tools, incorporating Apache Spark™ and Databricks notebooks for data processing, alongside Unity Catalog for efficient data management and governance. 


The DBRX models emerge as open, versatile Large Language Models (LLMs), engineered and licensed for application across both commercial and research endeavors. These models offer the flexibility to be fine-tuned for a wide range of domain-specific tasks in natural language processing and coding. 


DBRX Instruct enhances the base model through fine-tuning for concise instructional interactions. This model serves as a ready-to-use tool for few-turn question-answering, catering to queries in general English and programming tasks. 

Getting Started 

For those eager to explore these models, Microsoft Azure provides comprehensive documentation and guides to help you get started. Whether you're looking to enhance your application with natural language understanding or build an instructive AI agent, the combination of Databricks' models and Azure's platform offers a powerful toolkit for any developer's arsenal. 

Incorporating Databricks' databricks/dbrx-base and databricks/dbrx-instruct models into the Microsoft Azure AI model catalog represents not just a technical enhancement but a stride towards ensuring our dedication to providing a range of model choices to our customers.  As we delve into the capabilities of these models, we're optimistic about the prospects for developing more sophisticated, streamlined, and user-friendly AI applications. We invite the tech community to explore these tools and extend the limits of AI's potential. 

To learn more about the new models visit the Databricks Technical Blog and the Databricks Founder's blog. Try out the new dbrx models in Azure AI Studio model catalog. 


This article was originally published by Microsoft's AI - Machine Learning Blog. You can find the original article here.