MLflow
Open-source AI engineering platform: ML experiment tracking and model registry plus GenAI tracing, LLM evaluation, prompt registry and AI gateway
Overview
MLflow calls itself "the leading open source AI engineering platform." It started as the classic ML lifecycle tool (experiment tracking, model packaging and a model registry) and the 3.x line now adds a GenAI layer: OpenTelemetry-based tracing for any LLM or agent framework, more than 50 built-in evaluation metrics and LLM judges, a prompt registry with versioning and automatic prompt optimization, an OpenAI-compatible AI Gateway with routing, fallbacks and rate limits, and an Agent Server for deploying agents. Databricks launched it as open source in June 2018; it is Apache-2.0 licensed, reports over 30 million downloads a month and 900+ contributors, and the latest release is 3.18.0 (November 4, 2026). MLflow itself is free; managed hosting comes from Databricks, Amazon SageMaker and other clouds.
The Verdict
Who Should Use MLflow?
Best For
- Teams that want a free, vendor-neutral tracker and model registry they can self-host
- Databricks or SageMaker users, who get it as a managed service
- Teams running classic ML and LLM agents side by side on one platform
Not Ideal For
- Teams that want a polished hosted UI with no servers to run (W&B or Comet are simpler)
- Pipeline orchestration on its own (pair it with Metaflow, ZenML or Kubeflow)
What's Great
- Free and Apache-2.0, with the largest open-source community in this category
- One tool covers ML tracking, model registry, LLM tracing, evals and prompts
- Tracing is built on OpenTelemetry, so traces are not locked to one vendor
- Managed versions on Databricks and SageMaker remove the hosting work
Watch Out For
- Self-hosting a tracking server, database and artifact store is ongoing ops work
- Open-source server has thin access control compared with paid trackers
- Few public reviews: G2 shows not enough reviews for buyer insights
Pricing
View all features & details
Classic ML
- Experiment tracking of parameters, metrics and artifacts
- Model registry with versions and aliases
- Model packaging and serving
GenAI and Agents
- OpenTelemetry tracing for LLM and agent frameworks
- 50+ evaluation metrics and LLM judges
- Prompt registry with versioning and optimization
- AI Gateway (OpenAI-compatible, routing, fallbacks, rate limits)
- Agent Server for deploying agents
Real-World Usage
Adoption
- 30M+ package downloads per month
- 900+ contributors
- Frequent releases through 2026 (3.18.0 in November)
Origin
- Announced by Databricks as an open-source alpha in June 2018
- Built into Databricks and offered on Amazon SageMaker
How It Compares
| Feature | MLflow | Weights & Biases | Comet | Langfuse |
|---|---|---|---|---|
| Model | Open source, free | Hosted SaaS, enterprise self-host | Hosted SaaS plus open-source Opik | Open source plus cloud |
| Classic ML tracking | Yes, with registry | Yes, with registry and sweeps | Yes, with registry | No |
| LLM tracing and evals | Yes (OpenTelemetry) | Yes (Weave) | Yes (Opik) | Yes |
| Paid entry point | Managed via Databricks or SageMaker | Pro from $60 per month | Pro $19 per user per month | Cloud plans |
Sources: mlflow.org · GitHub · W&B pricing · Comet pricing
Is this your tool? Add a badge