What is MLflow?
MLflow started in 2018 as Databricks' open-source tool for machine learning bookkeeping: experiment tracking logs each training run's parameters, metrics and files, and the model registry versions the resulting models for deployment. Since the 3.x line it also covers LLM applications. Tracing records each model call, retrieval and tool call as spans, with one-line autologging for OpenAI, Anthropic, LangChain, LlamaIndex, DSPy and others, and accepts OpenTelemetry data. mlflow.genai.evaluate scores an app against a dataset with built-in LLM judges or your own scorers, a prompt registry keeps immutable prompt versions with movable aliases, and an AI Gateway routes requests across model providers.
The whole repository is under Apache-2.0, with no commercial-licence folder, and version 3.17.2 was released on 21 October 2026. The Python package needs Python 3.10 or later; TypeScript and Java clients exist for tracing.
Self-hosting is free. Managed MLflow is offered by Databricks, AWS SageMaker, Azure ML, Nebius and GCP. Databricks does not publish a separate price: MLflow comes with a workspace, which is billed for the compute it uses, and its Free Edition is for personal, non-commercial use only.
Is MLflow worth it?
MLflow is the sensible single system for a team that trains its own models and also ships LLM features: one Apache-2.0 server holds training runs, registered models, agent traces, evaluation results and prompt versions, and no feature in the code sits behind a paid licence. It has about 28,600 GitHub stars, released v3.17.2 on 21 October 2026, and its README claims more than 60 million monthly downloads.
The LLM half is now complete enough to stand on its own. Tracing is one line per library, such as mlflow.openai.autolog(), and the docs say the separate mlflow-tracing package has a 95% smaller footprint for production services. Evaluation runs a dataset through your app and grades it with built-in judges such as Correctness and Guidelines or with plain Python @scorer functions, and feedback from reviewers is stored on the trace itself. Against Langfuse, the operational load is lighter: MLflow starts on a SQLite file (the default since 3.7.0) and grows into PostgreSQL plus S3, GCS or Azure Blob storage, where Langfuse's production build needs ClickHouse, Postgres, Redis and object storage. Langfuse is the more focused LLM tool, with a TypeScript web app and published cloud pricing; MLflow's centre of gravity remains Python.
The weak point is security by default. mlflow server asks for no login unless it is started with --app-name basic-auth, and the 2026 advisories cluster on exactly that setup: in October a critical remote code execution and a high-severity arbitrary file read on unauthenticated servers, both fixed in 3.17.0. Older versions shipped the admin password password1234, which is now refused. Anonymous usage telemetry has been on by default since 3.2.0; it excludes parameters, file names and model content, and MLFLOW_DISABLE_TELEMETRY=true or DO_NOT_TRACK=true switches it off.
Upgrades also need care. Version 3.17.0 requires stopping every writer, taking a backup and running mlflow db upgrade, with no mixed-version rolling upgrade. That is the cost of owning the database, and also the point of it.
Weights & Biases remains the more polished hosted service for experiment dashboards, and its LLM layer, Weave, bills by data volume: Pro starts at $60 a month with 1.5 GB, then $0.10 per MB. MLflow costs the price of a small server and a Postgres instance, and its run history stays in tables you can back up and query yourself, which matters more to most teams than a nicer chart.
Where it earns its place
A four-person data science team at a retailer training demand-forecast models
Compare hundreds of training runs and promote the winning model to production
Each run logs parameters, metrics and the model file to a shared MLflow server; the registry versions models and an alias marks the one in production. The only cost is the server, a PostgreSQL database and an S3 bucket.
A platform engineer at a bank adding an LLM assistant to an existing ML stack
Trace and evaluate the assistant without buying a separate observability product
Autologging and the mlflow-tracing package send spans to the MLflow server the ML team already runs; mlflow.genai.evaluate grades releases against a test set. Before go-live, enable basic-auth or SSO and upgrade to 3.17.0 or later to close the October 2026 advisories.
An analytics lead at a company already on Databricks
Give analysts tracked experiments and governed models without running servers
Managed MLflow is built into the workspace and tied to Unity Catalog and Model Serving, with no separate MLflow fee; the bill is the Databricks compute the jobs use, which Databricks does not publish as an MLflow price.
Before you commit
- Setup
- uvx mlflow server starts a local server on port 5000 backed by SQLite; production setups use the Docker Compose bundle (MLflow, PostgreSQL and RustFS) or the official Helm chart.
- Integrations
- Python, TypeScript and Java SDKs; autologging for OpenAI, Anthropic, LangChain, LlamaIndex, DSPy, Pydantic AI and Vercel AI; OpenTelemetry ingestion and export; artifact storage on S3, GCS or Azure Blob; managed versions on Databricks, SageMaker, Azure ML, Nebius and GCP.
- Data and privacy
- Runs, traces and prompts stay in your own database and object store. Anonymous usage telemetry to mlflow-telemetry.io is on by default since 3.2.0 and excludes parameters, file names and model content; MLFLOW_DISABLE_TELEMETRY=true or DO_NOT_TRACK=true disables it. Nine advisories since July 2026, four rated high or critical.
- Cost at scale
- Self-hosting is the cost of one application server, a PostgreSQL database and object storage; the docs do not publish sizing. Databricks charges no separate MLflow fee and bills workspace compute instead; its Free Edition excludes commercial use.
- Limits
- About 2,225 open issues and pull requests. Basic-auth has no login-attempt limit and logout only by closing the browser; network protections such as --allowed-hosts work only on the default uvicorn server, not with --gunicorn-opts or --waitress-opts.
What’s good
- Experiment tracking, model registry, LLM tracing, evaluation and prompts in one Apache-2.0 server
- No commercial-licence features in the codebase
- Starts on SQLite; grows into PostgreSQL and S3, GCS or Azure Blob
- OpenTelemetry-compatible tracing with one-line autologging for major LLM libraries
What’s not
- Server runs without authentication unless basic-auth or SSO is configured
- Nine security advisories published since July 2026, including a critical RCE on unauthenticated servers
- Anonymous telemetry on by default
- Schema upgrades need downtime; no mixed-version rolling upgrades
Best for
Python teams that train models and build LLM features and want one free, self-hosted system of record for both.
MLflow alternatives
MLflow: common questions
- Is MLflow worth it?
- Still the default place to record ML experiments, now with LLM tracing and evaluation thrown in for free; lock the server down before anyone else can reach it. We rate it 4.4 out of 5.
- How much does MLflow cost?
- Free and open source under Apache-2.0; you pay only for the server, database and object storage. Managed MLflow on Databricks has no separate list price and is billed through Databricks compute; Free Edition is personal use only.
- What are the downsides of MLflow?
- Server runs without authentication unless basic-auth or SSO is configured; nine security advisories published since July 2026, including a critical RCE on unauthenticated servers; anonymous telemetry on by default; schema upgrades need downtime; no mixed-version rolling upgrades.
- Who is MLflow best for?
- Python teams that train models and build LLM features and want one free, self-hosted system of record for both.
- How much does MLflow cost for a team or in production?
- Self-hosting is the cost of one application server, a PostgreSQL database and object storage; the docs do not publish sizing. Databricks charges no separate MLflow fee and bills workspace compute instead; its Free Edition excludes commercial use.
- Is MLflow safe for company data?
- Runs, traces and prompts stay in your own database and object store. Anonymous usage telemetry to mlflow-telemetry.io is on by default since 3.2.0 and excludes parameters, file names and model content; MLFLOW_DISABLE_TELEMETRY=true or DO_NOT_TRACK=true disables it. Nine advisories since July 2026, four rated high or critical.
- What are the best MLflow alternatives?
- The closest alternatives we have reviewed are Langfuse, W&B Weave, Arize Phoenix and Opik.
Answers checked Nov 1, 2026.
Make MLflow? Show this rating on your site
