What is MLflow?
MLflow is an open-source platform that helps teams keep track of the many moving parts involved in building machine learning models and AI applications. It began with experiment tracking: a few lines added to a training script log parameters, metrics, code versions and output files to a tracking server, so data scientists can compare runs and reproduce results later. A model registry stores versioned models with aliases and metadata, and a standard packaging format lets the same model be served through a REST endpoint, a batch job or a cloud platform.
As large language models spread, MLflow grew into what its maintainers call an AI engineering platform. Tracing, built on OpenTelemetry, records each step an AI agent or application takes, including model calls, tool use and document retrieval. Evaluation tools score outputs with built-in metrics and LLM judges, a prompt registry keeps versions of prompts, and an AI gateway gives applications a single interface to multiple model providers with controls on access and cost. MLflow 3.0, released in June 2025, reorganized the project around these generative AI workflows while keeping its traditional machine learning features.
MLflow works with common frameworks including scikit-learn, PyTorch, TensorFlow, XGBoost, LangChain and the OpenAI SDK, and it can run on a laptop, a self-hosted server or a managed service. Databricks created the project and offers managed MLflow inside the Databricks platform, while Amazon SageMaker AI and other providers also host it. The project reports more than 30 million package downloads a month and more than 900 contributors.
Key features of MLflow
- 01
Experiment tracking
Runs log parameters, metrics, tags and artifacts such as plots and model files to a tracking server, where a web interface compares results across runs and experiments.
- 02
Model registry
A central registry stores versioned models with descriptions, tags and aliases, giving teams a shared record of which version has been approved and which one should be deployed.
- 03
Tracing for AI agents
OpenTelemetry-based tracing captures the inputs, outputs and latency of every step in an LLM application or agent, with automatic instrumentation available for many popular generative AI libraries.
- 04
Evaluation and LLM judges
Built-in metrics and configurable LLM judges score responses for qualities such as correctness, relevance and safety, and results can be compared across versions of an application.
- 05
Prompt registry
Prompts are stored with version history and can be linked to the traces and evaluations they produced, and optimization tools can propose improved prompts automatically.
- 06
AI gateway
A gateway provides one API in front of multiple large language model providers, with centralized credential management and controls on usage and cost.
- 07
Model packaging and serving
The MLflow Models format packages models from many frameworks together with their dependencies so they can be served as REST APIs, run in batch jobs or deployed to cloud platforms.
- 08
Framework autologging
Autologging integrations for libraries such as scikit-learn, PyTorch, XGBoost and LangChain capture metrics, parameters and traces without writing manual logging code.
Who uses MLflow?
- Data science teams record every training run so they can compare hyperparameters and metrics and reproduce the best model months later.
- Machine learning engineers manage model versions in a registry and hand approved versions to automated deployment pipelines.
- AI engineers trace and debug agents and retrieval-augmented generation apps to find which step produced a wrong or slow answer.
- Product teams evaluate chatbot and agent quality with LLM judges before releasing prompt or model changes to users.
History of MLflow
Databricks announced MLflow at its Spark + AI Summit in San Francisco in June 2018, with chief technology officer Matei Zaharia presenting it as an open-source way to track experiments, package code reproducibly and deploy models across different tools and clouds. Version 1.0 followed in June 2019. In June 2020 Databricks moved the project to the Linux Foundation to give it vendor-neutral governance; at the time MLflow had more than 200 contributors and over 2 million downloads a month.
MLflow 2.0 arrived in November 2022, and later 2.x releases added features for large language models, including an AI gateway and tracing. Amazon SageMaker launched a fully managed MLflow service in June 2024. MLflow 3.0, released in June 2025, made generative AI applications a central focus with expanded tracing, evaluation and version tracking, and the project had reached version 3.16 by September 2026.
- 2018
Databricks unveils MLflow as an open-source project at Spark + AI Summit.
- 2019
MLflow 1.0 is released in June.
- 2020
MLflow joins the Linux Foundation for vendor-neutral governance.
- 2022
MLflow 2.0 is released in November.
- 2024
Amazon SageMaker launches a fully managed MLflow service.
- 2025
MLflow 3.0 is released with a focus on generative AI tracing and evaluation.
MLflow pricing
Pricing modelFree and open source
Free under the Apache 2.0 license; managed MLflow is built into Databricks and offered by providers such as Amazon SageMaker AI, billed as part of those services.
Prices change often. Check the official MLflow website for current plans.
MLflow alternatives
Well-known alternativesWeights & Biases, Comet, Kubeflow, ClearML and Langfuse
MLflow: frequently asked questions
Who created MLflow?
MLflow was created by Databricks, the data and AI company founded by the original creators of Apache Spark, and was announced by Databricks CTO Matei Zaharia in June 2018. Since 2020 the project has been hosted by the Linux Foundation, and Databricks engineers develop it together with outside contributors.
What is MLflow used for?
MLflow is used to manage the lifecycle of machine learning models and AI applications. Teams use it to track experiments, compare training runs, register and deploy model versions, trace and debug AI agents, evaluate response quality with LLM judges and manage prompts.
Is MLflow free?
Yes. MLflow is free and open source under the Apache 2.0 license and can be self-hosted at no cost. Managed versions are available inside Databricks and on services such as Amazon SageMaker AI, where customers pay for the platform resources they use.
When was MLflow released?
Databricks released MLflow as an open-source alpha in June 2018. Version 1.0 followed in June 2019, version 2.0 in November 2022 and version 3.0, which focused on generative AI applications, in June 2025.
Is MLflow only for Databricks users?
No. MLflow is an independent open-source project that runs on any machine or cloud and works with many frameworks, including scikit-learn, PyTorch and LangChain. Databricks offers a managed version, but the same tracking and model formats can be used with a self-hosted server.
Sources
- MLflow mlflow.org
- The MLflow project joins Linux Foundation linuxfoundation.org
- Announcing MLflow 3 mlflow.org
- MLflow repository on GitHub github.com
Last reviewed . Spotted an error? Send a correction.