Which part of MLflow are you leaving: run tracking, data history, pipelines or the whole platform? Pick a hosted tracker if logging runs is the pain, a data versioning tool if nobody can rebuild last quarter's training set, a workflow framework if the hand-off to production is the problem, and a self-run platform if your data can't leave your own clusters.
MLflow has grown a long way past experiment logging. Its own site now calls it an open-source AI engineering platform, with LLM and agent tracing, evaluation, a prompt registry and an AI gateway sitting next to the classic tracking server and model registry. So when someone on your team says "let's replace MLflow", the first job is finding out which of those pieces they mean, because no single tool on this page replaces all of them in one move.
Name the part that hurts before you shortlist anything
Most teams use a slice of MLflow, not the lot. Write down which of these four jobs it does for you today, and which one is causing the complaint.
Experiment tracking is the obvious one: parameters, metrics, artifacts and the question "which run produced the model in production?". If your tracking server is slow, self-hosted by someone who has left, or nobody outside the ML team can read it, a hosted tracker is the shortest route out.
Data versioning is the one people forget they need until an audit or a bad retrain. MLflow logs what you tell it to, but it isn't built to give a dataset the same commit history as code. A Git-style data tool fixes that, and it usually sits beside a tracker rather than replacing one.
Pipelines and workflows cover the steps between a notebook and something running on a schedule. If your real problem is that every model reaches production by a different hand-built route, a workflow framework or orchestrator matters more than a nicer chart of loss curves.
Platforms on your own infrastructure are for teams where data location is the deciding rule. Kubernetes shops, regulated teams and anyone training on their own GPUs often want scheduling, tracking and governance in one control plane they run themselves.
Two rivals changed shape this year
Before you trial anything, check that the product you remember is the product you'll be buying. The best-known hosted tracker now has its documentation inside CoreWeave's docs, under the Forge name. One pipeline tool here, ZenML, now introduces itself on its homepage as ZenML Labs. The software is still there; what may have moved is who holds your contract, your account and your data.
For a tool that will hold years of run history, send the vendor three questions: which legal entity holds your data, where it's stored, and how you'd export it if you left. Get the answer in writing and keep it with the procurement notes.
Seven tools worth a look
One line each on what kind of tool it is. The guide around them does the choosing.
ClearML
A broad AI infrastructure platform bundling a compute control plane, a development centre and a GenAI app engine, for teams that want one system rather than several.
DVC
Git-style version control for datasets and models, so the data behind a run gets commits and branches just like the code that trained it.
Kubeflow
The Kubernetes-native choice: a CNCF graduated, open-source set of modular AI projects that cluster-running platform teams can adopt piece by piece.
Metaflow
A Python framework for building, versioning and shipping ML and data science workflows, with the move to production pitched as one command.
Polyaxon
A control plane you host yourself, scheduling, tracking and governing GPU training and agent sessions while the data stays on your own clusters.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
Weights & Biases
Hosted experiment tracking with sweeps, artifacts and reports, plus Weave for evaluating LLM and agent work; its documentation now lives under CoreWeave.
ZenML
Pipeline orchestration that runs on whichever infrastructure you pick, now trading as ZenML Labs and adding evals drawn from agent traces.
Getting your history out of MLflow
Most migration plans skip this step and pay for it later. MLflow's own GitHub organisation publishes a package called MLflow Export Import that copies runs, experiments and registered models from one tracking server to another. It works through the MLflow REST API, writes each object to an intermediate folder as JSON plus artifacts, and runs in single mode (one run, one experiment, one model) or bulk mode (a collection, up to a whole server). Its quick start exports an experiment with export-experiment --experiment sklearn-wine --output-dir /tmp/export.
That folder is your insurance policy, whatever you move to. The target tool won't read it directly, but it gives you the parameters, metrics and tags of every run in a format you can script against, and a copy that survives the old server being switched off. For artifacts on their own, the MLflow command line has mlflow artifacts download --run-id, which pulls a run's files into a local folder; it's the quickest way to rescue the model files behind the handful of runs that are in production today.
Three habits keep the move short. Export the lot, then migrate only the recent runs people still compare against, and keep the old server running read-only until nobody has asked about it for a few months. List every deployment script that loads a model by name from the MLflow registry, because each one breaks on the day you switch. And don't choose on charts: every tracker draws a decent loss curve, so judge permissions, artifact storage, search across tens of thousands of runs, and whether your LLM tracing fits the same tool or needs its own.
Run both side by side for a fortnight
No comparison page, this one included, can tell you how a tool feels on your own runs. This can, and the first part fits in an hour:
- Pick one live project with a model in production and a few weeks of recent runs.
- Add the candidate's logging call next to your existing MLflow call, so both receive the same parameters and metrics. That's usually a few lines.
- Run one real training job and check that the numbers match in both places.
- Export that run, with its artifacts, back out of the new tool. If the way out is painful in week one, it won't improve in year three.
- Leave both logging for two weeks, then ask someone who reads results, not just someone who logs them, to answer a real question from the new tool alone.
Hosted or self-run, and what each really costs
This is the fork that decides most shortlists. On their own sites, three of the seven show no paid plan at all: the Python workflow framework, the Kubernetes project and the data versioning tool, which points enterprise-scale needs to a sister product with no price shown. The other four pair a free plan or a self-hosted option with paid plans. Self-run tools cost nothing to download and a fair amount to keep alive: someone patches them, backs them up and answers the 2am message when the tracking database fills a disk. Hosted tools move that work to the vendor and the bill to you.
Where prices are published, read them on the vendors' own pricing pages and note what's charged per user, per team and for usage. The bill grows with seats, compute and stored artifacts, so the pricing page figure is a floor, not a forecast. For self-run tools, cost the engineer's time, not the licence: a steady slice of someone's week is the real price, and it rarely makes the comparison spreadsheet.
A rough rule: if you already run Kubernetes well and have a platform engineer who enjoys it, self-run is cheaper and keeps data where you want it. If your ML team is small and would rather train models than babysit servers, hosted wins, even when the invoice stings. The second fork is suite versus single job. A suite gives you one login and one bill; a data versioning layer paired with a workflow framework is easier to swap later, and for many teams two focused tools is the sensible end state, not a planning failure.
Where I'd start, by the part that hurts
Tracking is the pain and you want it off your hands: go hosted. Retrains can't be reproduced: add a data versioning tool first and keep MLflow for now, because that fixes the problem without a migration. Models reach production by hand: a workflow framework or an orchestrator is the fix, not a new tracker. Data must stay on your clusters, or you're already on Kubernetes: a self-run platform, with the engineer time written into the budget.
If your pipelines also need a scheduler, my list of Airflow alternatives covers that layer, and these Weights and Biases alternatives go wider on hosted tracking.
Do you make an experiment tracking, MLOps or ML workflow tool? Tool makers can ask to be considered for a featured place on this list.
Want your tool on this list? Featured spots and reviews here are paid partnerships. Here's how to partner with me.
