What is the difference between experiment tracking and pipeline orchestration, and which do you need first?
Teams use the word MLOps for two jobs that are often bought separately. Experiment tracking records parameters, metrics and artifacts so that you can compare runs afterwards. In this list, DagsHub and the managed MLflow service on Nebius AI Cloud are the clearest examples. Pipeline orchestration decides how code runs and in what order: ZenML wraps Python steps into pipelines, and Valohai executes and versions each job on the machines you point it at.
If your pain is that nobody can say which dataset and which parameters produced the model in production, start with lineage, which is what Valohai versions per execution and what DagsHub shows from model back to source data.
If your pain is that a notebook that works on a laptop cannot be repeated on a cluster, start with orchestration, where ZenML is the open source entry point. Teams that have both problems usually end up with two tools, and the vendors say so themselves: ZenML and several others point to DagsHub or MLflow for dense run charts.
The categories at the bottom of the ranking solve narrower problems and make sense only after the basics exist. Giskard tests models before release, Lightly picks which images to label, Flower trains without centralising data, and UbiOps serves finished models. None of them tracks experiments.
How do European MLOps platforms charge, and what can you actually compare?
Only a few vendors publish prices, and they publish different units. ZenML lists Scale SaaS at $999 per month with 2,000 executions, 3 projects and 5 snapshots, and its separate Kitaru Cloud product for agent sessions at a flat $39 per month.
DagsHub lists Team at $119 per user per month, or $99 per user billed yearly, with Individual at $0. Flower prices SuperGrid in euros per month billed yearly, with monthly credits and 3,000 sign-up credits, and Nebius publishes GPU hours, for example HGX H100 from $4.50 on demand.
Valohai charges a fixed fee per user with unlimited projects, experiments, pipelines and deployments, but the figure is on request. Hopsworks has a free tier of one project without a card, then usage-based SaaS with no published rates. KNIME says Business Hub plans scale with capacity and have no per-seat licensing, again without figures. UbiOps is pay per use in its cloud and quoted for Private.
The practical consequence is that a per-seat tool and a per-execution tool cannot be put in one spreadsheet without your own usage numbers. Count your active users, your pipeline runs per month and your tracked experiments first. Then ask the three vendors with unpublished prices for a quote on those same numbers, and keep in mind that three of the published prices are in US dollars.
Does choosing a European vendor keep training data out of US jurisdiction?
It changes the contract, not necessarily the servers. A contract with ZenML GmbH is governed by German law, one with Hopsworks AB by Swedish law, and Valohai Oy names the Finnish Data Protection Ombudsman as supervisory authority. That gives you a European counterparty and a European regulator, which is most of what a data protection officer asks for.
The infrastructure underneath is a separate question, and the tools answer it differently. Valohai, UbiOps Private, Hopsworks Enterprise and DagsHub Enterprise can run on-premises or air-gapped, which removes the hosting question entirely. ZenML, Giskard and Lightly publish open source code you can run yourself.
Flower keeps raw data with each participant, although the SuperGrid control plane is hosted by Flower Labs. Nebius lets you choose the region per cluster, and its privacy policy lists Nebius Inc. in the United States and Nebius Israel Ltd, so read which entity processes what.
Two entries are not EU entities. DagsHub is Israeli, and KNIME and Lightly are Swiss, a country with an EU adequacy decision but outside the EEA. Whether that matters depends on your policy, not on the tools.
How do you move off Weights & Biases without losing your run history?
Weights & Biases Models is now sold as part of CoreWeave Forge, and its privacy policy is issued by CoreWeave, Inc. of Livingston, New Jersey; on 23 October 2026 the pricing page listed a free tier, Pro from $60 per month for early-stage teams under 50 employees, and custom Enterprise. Teams that want a European counterparty usually start by freezing what they have: export the run history and artifacts you still need before the contract ends.
The easiest landing is a tracker that speaks MLflow. DagsHub describes its tracking as MLflow compatible, and Nebius runs standard MLflow, so logging code written against the MLflow API carries over with a changed server address. Moving the training code onto ZenML or Valohai takes more work, because Valohai expects files under /valohai/inputs and /valohai/outputs, parameters through argparse and metrics printed as JSON, and ZenML wants steps written as Python functions.
Do not expect a one-to-one dashboard match. ZenML states plainly that it records and orchestrates rather than offering dense run comparison charts, and Valohai is described as lighter on dashboards than a dedicated tracker. Run a real project on the new tool for a few weeks before cancelling the old one.
When should you not switch to a European MLOps platform?
When the tracking server is not your actual exposure. If your runs contain only public benchmark data and hyperparameters, a jurisdiction argument adds little, and the free tier of the incumbent may be cheaper than the migration. The same holds if your team depends on features no European tool here replicates, such as the run-comparison dashboards that ZenML, Valohai and UbiOps openly do not try to match.
It also holds if you need a single product. This category is fragmented: Hopsworks has a feature store but tracking is a smaller part of it, KNIME has no dedicated experiment tracker, and Nebius is a GPU cloud with MLflow attached. Assembling two or three tools means more contracts, more integration and more people who need to know them.
Finally, check the maturity of the paperwork. DagsHub's terms were last updated in January 2022, Valohai's public terms are version 8 from 2021, and UbiOps's General Terms were last modified in October 2023. None of that is disqualifying, but ask for the current contract before you rely on a public page.
How do the specialist tools fit next to a main platform?
Four of the ten tools solve one step and expect a partner. Giskard generates security and quality tests for models and agents (prompt injection, data disclosure, hallucination) and plugs into CI/CD through an SDK, so it sits after training and before release. Lightly selects which images are worth labeling using embedding-based sampling, which sits before training. Flower replaces the data-gathering step itself when the data cannot be copied. Hopsworks is the data layer that training and serving both read.
A sensible stack for a vision team in Europe might be Lightly for the dataset, DagsHub or Nebius MLflow for tracking, ZenML for pipelines and UbiOps for serving, with Giskard as a release gate. That is four vendors in four countries, which is a legitimate outcome and also the argument for choosing fewer, wider tools such as Valohai or Hopsworks.
Whatever you combine, check the integration points the vendors name: DagsHub and Nebius speak MLflow, Hopsworks works with Iceberg, Delta and Hudi tables, and Lightly exports COCO, Pascal VOC and KITTI.
