MLflow is open source. Databricks' MLflow documentation, last updated 6 October 2026, describes it as "the largest open source AI engineering platform for agents, LLMs, and ML models", with over 30 million monthly downloads. Databricks created it, released it as open source in June 2018, and the project is now backed by the Linux Foundation with the code on GitHub under the Apache licence.
The part that confuses people is that two things share the name. There is MLflow the open source project, which you can pip install and run against your own tracking server, and there is Databricks-managed MLflow, which the documentation calls "a fully managed and hosted version of MLflow, building on the open source experience to make it more reliable and scalable for enterprise use". Managed MLflow went to public preview in March 2019 and generally available on 25 April 2019. Both are MLflow; one of them is a product with a bill attached.
What the open source version gives you #
The project covers experiment tracking, model evaluation, a production model registry and model deployment tools for traditional ML, plus observability, evaluation, prompt management and an AI Gateway for agents and LLM applications. It supports any LLM provider, agent framework and ML library, with native SDKs for Python, TypeScript and JavaScript, Java and R. MLflow 3 added Logged Models, which persist across environments and runs and carry links to metadata, metrics, parameters and the code that produced the model, and deployment jobs for evaluation, approval and deployment steps.
None of that requires Databricks. A team running MLflow on its own infrastructure gets the tracking, the registry and the tracing.
What the Databricks version adds #
The managed version is built on Unity Catalog. Model Registry is integrated with Unity Catalog, deployment job workflows are governed by Unity Catalog, and every event lands in an activity log on the model version page. Traces are stored in Unity Catalog as Delta tables, which the documentation recommends, and that gives governed access, SQL queryability and no per-experiment storage cap, the same governance model as any other Unity Catalog table. You can query those traces with SQL or ask about them in natural language through Genie Code.
The rest of the platform joins up too: automated feature lookups from Feature Store, Model Serving to a REST API endpoint, and request and response capture for monitoring, augmented with trace data per request.
The licence question behind the question #
Teams usually ask whether MLflow is open source because a procurement or architecture review needs to know what happens if they leave Databricks. Experiment code, logged models and the tracking APIs are the open project, so they travel. The Unity Catalog governance, the Delta-backed trace storage and the serving integration are Databricks, and those do not. MLflow and the Model Lifecycle goes through the four things MLflow does and where the Databricks version diverges, and Agents on Databricks covers evaluation, which is where most teams meet MLflow 3 first.
Questions people ask #
Is MLflow open source? #
Yes. Databricks introduced MLflow as an open source machine learning platform in June 2018, and its documentation describes MLflow as the largest open source AI engineering platform for agents, LLMs and ML models, licensed under Apache and backed by the Linux Foundation.
Is MLflow free to use? #
The open source project is free to download and run yourself, with over 30 million downloads a month. Databricks-managed MLflow is a hosted commercial service on the Databricks platform, and it has been generally available since 25 April 2019.
What is the difference between MLflow and managed MLflow on Databricks? #
Managed MLflow is the same project hosted by Databricks and built on Unity Catalog, so the Model Registry, deployment job workflows and trace storage are governed there. Traces are stored as Unity Catalog Delta tables with SQL queryability and no per-experiment storage cap.