Databricks Mosaic AI
Build, serve, and govern models and agents directly on the Databricks lakehouse.
Mosaic AI is what Databricks turned the $1.3 billion MosaicML acquisition into: model training, serving, vector search, and agent tooling wired into Unity Catalog governance. Its Agent Bricks layer — over 100,000 agents built since launch, with Databricks reporting agents processing more than a quadrillion tokens a year — auto-tunes task-specific agents against your own data. Foundation Model APIs serve open and frontier models pay-per-token, so teams can prototype without provisioning a single GPU.
Best for
Data and ML platform teams already running Databricks who want training, serving, evaluation, and agent development inside the same governance boundary as their data. It rewards committed platform users far more than teams shopping for a standalone model endpoint.
Snowflake Cortex bolts AI functions onto a warehouse; Mosaic AI goes deeper down the stack — you can pretrain or fine-tune your own model, not just call someone else's. Against AWS SageMaker and Google Vertex AI, its edge is Unity Catalog: lineage, permissions, and evaluation live where the data lives. The Tecton acquisition added real-time feature serving that raw cloud endpoints don't match.
Core features
Model workflows
Pretraining, fine-tuning, and serving for open and proprietary models, plus pay-per-token Foundation Model APIs and vector search for RAG pipelines.
Data governance
Unity Catalog extends table-level lineage, permissions, and audit to models, features, and agent tools — the capability Snowflake and bare cloud endpoints compete hardest against.
Enterprise deployment
Provisioned-throughput serving, MLflow-based evaluation, real-time feature serving via the Tecton acquisition, and Agent Bricks for production agents on governed data.
Pros
- + Training, serving, vector search, evaluation, and agent building share one governance layer, so model lineage is auditable the same way table lineage is
- + Agent Bricks automates the tuning-and-eval loop — over 100,000 agents built since launch, per Databricks
- + Pay-per-token Foundation Model APIs let teams test open and frontier models without provisioning a single GPU
Watchouts
- - Real cost is DBUs plus your own cloud compute; all-in bills commonly run 50–100% above the headline DBU rate
- - Only worth adopting with a genuine Databricks platform commitment — as a standalone AI stack it loses to simpler, cheaper endpoints
Alternatives
Frequently asked
What is Databricks Mosaic AI?
Build, serve, and govern models and agents directly on the Databricks lakehouse.
Who builds Databricks Mosaic AI?
Databricks Mosaic AI is built by Databricks.
How much does Databricks Mosaic AI cost?
Databricks Mosaic AI uses usage based. Current tiers: Pay-per-token model APIs · Mosaic AI compute from ~$0.07/DBU. Check the official site before buying — AI pricing changes often.