Overview

Databricks Mosaic AI is a component of the Databricks Lakehouse Platform designed to support the development and deployment of generative AI applications and large language models (LLMs). It integrates with the existing data management and processing capabilities of Databricks, providing a unified environment for the entire machine learning lifecycle, from data ingestion and preparation to model training, fine-tuning, and production deployment Databricks Mosaic AI product overview. The platform targets developers and technical buyers who require a scalable and governed solution for building custom AI applications leveraging proprietary data.

The core offerings within Mosaic AI address key challenges in enterprise AI, including model governance, real-time serving, and the integration of open-source and proprietary models. For instance, it provides tools for fine-tuning LLMs on custom datasets, allowing enterprises to adapt general-purpose models to specific business contexts while maintaining data privacy and security. This is particularly relevant for applications that require domain-specific knowledge or adhere to strict regulatory compliance Mosaic AI capabilities.

Mosaic AI is engineered for production-ready deployments, emphasizing MLOps practices. It includes features for monitoring model performance, managing model versions, and ensuring the reliability of AI systems in real-time use cases. The platform's integration with Unity Catalog provides a centralized metadata layer, enhancing data and AI governance across an organization Databricks Machine Learning documentation. This approach aims to streamline the transition of AI projects from experimentation to operational use, reducing the complexity associated with scaling AI initiatives.

Key features

  • Mosaic AI Model Serving: Provides scalable, real-time inference for large language models and other ML models, supporting both open-source and proprietary models Databricks Model Serving.
  • Mosaic AI Vector Search: Offers a vector database optimized for Retrieval Augmented Generation (RAG) applications, enabling LLMs to access and incorporate up-to-date, proprietary data Mosaic AI Vector Search capabilities.
  • Mosaic AI Playground: A web-based interface for experimenting with and testing LLMs, facilitating rapid prototyping and development of generative AI applications.
  • Mosaic AI Agent Framework: Tools and libraries for building AI agents that can interact with other systems and perform complex tasks, often leveraging LLMs.
  • Mosaic AI Gateway: Centralized management and routing for LLM APIs, providing security, rate limiting, and observability across various models Mosaic AI Gateway features.
  • Unity Catalog: A unified metadata and governance layer for data, analytics, and AI, ensuring data quality, access control, and lineage across the Databricks Lakehouse Platform Unity Catalog documentation.
  • MLflow Integration: Fully integrated with MLflow for experiment tracking, model lifecycle management, and reproducible AI workflows MLflow on Databricks.
  • LLM Fine-tuning: Capabilities to fine-tune pre-trained LLMs on custom datasets to improve performance on specific tasks or domains.

Pricing

Databricks Mosaic AI is part of the broader Databricks Lakehouse Platform, which offers custom enterprise pricing. Costs are primarily based on usage, including compute resources (DBUs - Databricks Units) consumed, data storage, and specific service usage like model serving endpoints Databricks Pricing Page. Customers generally engage in direct discussions with the Databricks sales team for detailed quotes tailored to their specific consumption patterns and architectural requirements.

Service Component Pricing Model Notes
Databricks Units (DBUs) Consumption-based Compute for notebooks, jobs, SQL warehouses, and ML workloads. Varies by cloud provider and region.
Data Storage Per GB per month Storage in Unity Catalog and other associated data services.
Mosaic AI Model Serving Per token/inference, per endpoint-hour Specific pricing for real-time model serving endpoints and LLM inference.
Mosaic AI Vector Search Per vector index unit, per query Costs associated with maintaining vector indexes and performing similarity searches.
Unity Catalog Included with DBU usage, some storage costs Governance features are generally part of the platform's overall cost structure.

Pricing as of June 2026. Custom enterprise pricing is typical, requiring direct contact with Databricks sales.

Common integrations

Alternatives

  • AWS SageMaker: A cloud machine learning service that provides tools for building, training, and deploying ML models, including capabilities for LLMs AWS SageMaker product page.
  • Google Cloud Vertex AI: A managed ML platform that offers a unified environment for building and deploying ML models, with extensive support for generative AI and MLOps Google Cloud Vertex AI overview.
  • Azure Machine Learning: Microsoft's cloud-based platform for end-to-end machine learning, providing services for model development, training, deployment, and management Azure Machine Learning product details.
  • H2O.ai: Offers an AI Cloud platform for automated machine learning and MLOps, including tools for building and deploying AI applications H2O.ai homepage.

Getting started

To begin with Databricks Mosaic AI, users typically start within a Databricks Workspace. The following Python code snippet demonstrates how to interact with a model served via Mosaic AI, for example, an LLM for text generation. This example assumes a model endpoint has already been created and is accessible.

from databricks.sdk import WorkspaceClient
from databricks.sdk.service.serving import QueryEndpointInput

# Initialize the Databricks Workspace Client
w = WorkspaceClient()

# Define the name of your Mosaic AI Model Serving endpoint
endpoint_name = "your-llm-serving-endpoint"

# Define the input payload for the LLM
# This structure may vary based on the specific model deployed
input_data = {
    "dataframe_split": {
        "columns": ["prompt"],
        "data": [["What is the capital of France?"]]
    }
}

# Query the model serving endpoint
response = w.serving_endpoints.query_endpoint(
    name=endpoint_name,
    query_endpoint_input=QueryEndpointInput(**input_data)
)

# Print the response
print(response.as_dict())

This example uses the Databricks SDK for Python to interact with a deployed model serving endpoint. Users would replace "your-llm-serving-endpoint" with the actual name of their deployed model and adjust the input_data structure to match the expected input format of their specific LLM Databricks LLM Serving documentation.