European Alternatives to Weights & Biases
Looking for a European alternative to Weights & Biases? Weights & Biases Models is now sold as part of CoreWeave Forge, and the privacy policy that covers the platform is issued by CoreWeave, Inc., a company based in Livingston, New Jersey. On 23 October 2026 its pricing page listed a free tier, a Pro plan from $60 per month for early-stage teams under 50 employees, and custom Enterprise plans.
The European tools below track experiments, version data and run pipelines under contracting entities in Finland, Germany, Sweden, the Netherlands, France, Switzerland and Israel, and several can be self-hosted.
How we rank these tools — 4-step process
-
1
European ownership, verified
The company is headquartered and incorporated in the EU, EEA, the UK, Switzerland or Israel, and processes customer data in Europe or Israel. A US parent company disqualifies a tool from this page regardless of where its servers are.
-
2
Category fit and hands-on review
What the tool actually does, who it suits, and where it falls short — checked against the vendor’s own documentation, changelog and pricing page rather than its marketing copy.
-
3
Compliance and pricing check
GDPR posture, hosting location and the prices quoted on this page are verified against the vendor’s public pricing before publication, and re-checked when we revisit the category.
-
4
Position on this page
Placement on this page can be paid, and that can affect which tools appear here and the order they appear in. It never buys a good review: a tool that fails the checks above is not here at any price, and payment does not change the shortcomings we write about. A vendor can ask us to correct a factual error — not to remove a criticism.
Vendors can pay for visibility on this page. It never changes what an entry says about a product, including the criticism, and we earn nothing when you click through to a vendor. Paid placement can affect which tools appear here and the order they appear in. Editorial policy
10 European Alternatives to Weights & Biases
ZenML
Munich open source pipeline orchestration with artifact and model tracking
Valohai
Turku platform for versioned ML runs, pipelines and GPU scheduling
Hopsworks
Stockholm feature store with model registry and serving, free tier
DagsHub
Experiment tracking, data versioning and annotation on one platform
UbiOps
Hague platform to deploy and orchestrate models on any infrastructure
KNIME
Zurich visual workflows with a Business Hub for deploying models
Giskard
Open-core testing and red-teaming for ML models and AI agents
Lightly
Zurich data curation and labeling for computer vision training sets
Flower
Hamburg federated learning framework with a hosted SuperGrid service
Nebius AI Cloud
Amsterdam GPU cloud with managed MLflow experiment tracking
Key takeaways
- Weights & Biases Models is sold as part of CoreWeave Forge, and its platform privacy policy is issued by CoreWeave, Inc. of Livingston, New Jersey.
- On 23 October 2026 the W&B pricing page listed a free tier, Pro from $60 per month for teams under 50 employees, and custom Enterprise plans.
- DagsHub and Nebius AI Cloud are the closest European tracking options, both usable with standard MLflow logging code.
- ZenML (Munich) and Valohai (Turku) replace the pipeline and run-versioning side, and neither copies the Weights & Biases dashboards.
- Of the ten alternatives, DagsHub is Israeli and KNIME and Lightly are Swiss, so not every entry is an EU entity.
Why people leave Weights & Biases
Weights & Biases became the default experiment tracker for many machine learning teams because logging a run took a few lines of code and the dashboards for comparing runs were quick to read. A researcher could share a link, and a colleague saw the same charts. That habit is why the product shows up in so many training scripts, and why switching it feels bigger than switching an ordinary SaaS tool.
What has changed is who stands behind it. Weights & Biases Models is now sold as part of CoreWeave Forge, and wandb.ai/site/pricing redirects to a CoreWeave pricing page. The privacy policy that covers the platform is issued by CoreWeave, Inc., a company based in Livingston, New Jersey. For a European buyer that means the party processing run metadata, dataset names and artifact lineage is an American company, under American law, with a GPU cloud business next to the tracker.
We could not read the W&B terms page itself, because it renders only in the browser, so we do not describe its contract terms here. What we can say is narrower: the tracking server for a model you train on customer data is operated by a US group, and a European alternative gives you a different counterparty and a different regulator to deal with.
- The vendor is now part of a US infrastructure company Models sits inside CoreWeave Forge, and CoreWeave, Inc. issues the privacy policy for the platform. A tracking tool that belonged to a standalone software company is now a product of a group that also sells compute. Whether that matters depends on your policy, but a procurement team that approved one vendor in 2022 may find a different one on the invoice today.
- Pricing is aimed at small, early-stage teams first On 23 October 2026 the pricing page listed a free tier, a Pro plan from $60 per month for early-stage teams under 50 employees, and custom Enterprise plans. A company past that size lands in a quote conversation. The European tools here show different models, from ZenML's free open source core and DagsHub's $99 per user billed yearly to Valohai's per-user licence on request.
- Self-hosting and sovereignty options are Enterprise matters The pricing page we read points larger customers to Enterprise. Several European tools offer a self-hosted or air-gapped route as a normal part of the product: Valohai runs on-premises and air-gapped, UbiOps Private installs in your environment, and ZenML, Giskard and LightlyStudio publish open source code you can run yourself.
- US jurisdiction over the party that holds your run history Run history can contain dataset names, feature lists, prompts and evaluation sets. With a US vendor those fall under US law and US government access rules, whatever region the servers are in. A German, Finnish, Swedish or Dutch contracting entity moves the dispute and the supervisory authority back to Europe, which is the point of this directory.
What you have to replace, not just match
Weights & Biases is not a single feature. Most teams use it for four things at once: logging metrics and parameters from training code, comparing runs in charts, storing artifacts such as datasets and model files with lineage, and sharing reports with colleagues. Work out which of these you really depend on, because no European tool here covers all four in one product.
Logging and comparison map to DagsHub (MLflow compatible) or to managed MLflow on Nebius AI Cloud. Artifacts and lineage map to ZenML, which stores them for each pipeline run, and to Valohai, which versions every execution. If you used W&B Tables to curate image data, Lightly covers that for vision. If you evaluated language-model behaviour, Giskard covers testing.
What you will not find is a drop-in clone of the dashboards. ZenML says outright that it records and orchestrates rather than offering dense run-comparison charts. Plan for the training code to change too: Valohai expects a specific file and parameter layout, and ZenML expects steps as Python functions.
The alternatives compared
| Position | Tool | Headquarters | Pricing | Jurisdiction |
|---|---|---|---|---|
| #1 | ZenML | Munich, Germany | Open source free; Scale SaaS $999/month; Enterprise on request | EU (Germany) |
| #2 | Valohai | Turku, Finland | Per-user licence, quote on request; unlimited projects, experiments, pipelines and deployments | EU (Finland) |
| #3 | Hopsworks | Stockholm, Sweden | Free tier (1 project, no card); SaaS pay as you go; Enterprise on request | EU (Sweden) |
| #4 | DagsHub | Israel | Individual $0; Team $119/user/month ($99 billed yearly); Enterprise custom | Israel (outside the EU/EEA) |
| #5 | UbiOps | The Hague, Netherlands | UbiOps Cloud pay per use; UbiOps Private on request; no prices published | EU (Netherlands) |
| #6 | KNIME | Zurich, Switzerland | Analytics Platform free; Business Hub priced by capacity, no per-seat licensing, on request | Switzerland (adequacy decision, outside the EEA) |
| #7 | Giskard | France | Open source free; Professional on request | EU (France) |
| #8 | Lightly | Zurich, Switzerland | LightlyStudio free to try; commercial pricing on request via demo | Switzerland (adequacy decision, outside the EEA) |
| #9 | Flower | Hamburg, Germany | SuperGrid paid plans in euros per month billed yearly with monthly credits; 3,000 sign-up credits; Max on request | EU (Germany) |
| #10 | Nebius AI Cloud | Amsterdam, Netherlands | GPU prices per hour on its price list, for example HGX H100 from $4.50 on demand; MLflow cost see website | EU (Netherlands) |
How each alternative compares to Weights & Biases
Best for: Engineers who want repeatable pipelines and artifact lineage from a German vendor
ZenML GmbH of Munich (HRB 268487) answers a different question from Weights & Biases. W&B records what happened during a run; ZenML defines the run. You write steps in Python, ZenML executes them on the orchestrator and cloud you choose, and each execution stores its artifacts and metadata, so a model can be traced to the data and code that made it. That lineage is the part of W&B that many teams use most after the charts.
Against the American incumbent the case is licence and jurisdiction. The core is Apache 2.0, the terms are governed by German law, and the home page lists SOC 2 Type II and ISO 27001. The honest gap is the chart view: ZenML does not ship dense run-comparison dashboards, and Scale SaaS at $999 per month is a large step up from free.
What ZenML does better than Weights & Biases
- Apache 2.0 core you can host yourself, where W&B Enterprise is the sold route to control
- Pipelines that move between orchestrators and clouds without a rewrite
- German contracting entity with German governing law instead of a US group
- Free plan without an execution cap, against a free W&B tier with plan limits
Where ZenML is a step down from Weights & Biases
- No chart-heavy run comparison dashboard on the plans listed
- Basic model registry in the open source edition
- Scale SaaS at $999 per month, in dollars
Standout against Weights & Biases. It treats the pipeline, not the dashboard, as the thing worth versioning, and gives that away under Apache 2.0.
Best for: Teams that own GPU clusters or several clouds and need every execution reproducible
Valohai Oy of Turku has run since 2016 and sells a platform that executes training jobs rather than watching them. Code, data, parameters and environment are recorded for every run, pipelines cache each step, and the vendor describes one-click reproduction of an earlier run. Jobs go to virtual machines, Kubernetes or Slurm, in a cloud or your own data centre, including air-gapped sites.
That is a larger commitment than adding a tracking call to a script. Your code has to follow Valohai layout rules, and the per-user price is not published, so you cannot compare it with the W&B Pro plan without a quote. In return you get lineage from deployment back to dataset, and a Finnish supervisory authority named in the privacy policy.
What Valohai does better than Weights & Biases
- Runs the jobs itself, so reproduction does not depend on discipline in logging code
- On-premises and air-gapped deployment as a standard option
- Unlimited projects, experiments and pipelines under one per-user fee
- Finnish entity with a published Turku address
Where Valohai is a step down from Weights & Biases
- Price per user is not published
- Lighter team dashboards than a dedicated tracking tool
- Code must follow Valohai file and metric conventions
Standout against Weights & Biases. Every run is traced to the exact dataset and preprocessing step, and an approval gate can block a model that does not beat its baseline.
Best for: Teams whose models depend on shared, consistent features in training and production
Hopsworks AB of Stockholm sells the data side of MLOps. A feature store keeps the same feature values available to training and to online serving, with a model registry and serving on top, and the vendor cites sub-millisecond lookups using RonDB. It works with Iceberg, Delta and Hudi tables, and Zalando is quoted as a production user.
Against Weights & Biases this is a complement more than a replacement. Experiment tracking is a smaller part of the product, so a team that lives in run charts will probably still want DagsHub or MLflow. What Hopsworks adds is a free tier of one project without a card, a Swedish contracting entity, and on-premises or air-gapped operation on Enterprise.
What Hopsworks does better than Weights & Biases
- Feature storage and serving, which sit outside the W&B tracking product
- Free tier with a project, feature store and registry, no card needed
- On-premises and air-gapped operation on Enterprise
- Swedish company under Swedish law
Where Hopsworks is a step down from Weights & Biases
- Experiment tracking is a smaller part of the product
- No published SaaS unit prices
- Self-hosting suggests a four to five node Kubernetes cluster
Standout against Weights & Biases. It is the only European entry built around a feature store, with serving and a registry in the same platform.
Best for: Small and mid-sized teams that want MLflow-compatible tracking plus data versioning
DAGsHub Ltd. is governed by Israeli law, which counts as European for this directory since 22 October 2026 but is not an EU or EEA entity, and buyers with a strict EU rule should note that. Functionally it is the closest to what W&B users do day to day: tracking described as MLflow compatible, data and model versioning, lineage from model to source data, notebook diffing and Label Studio compatible annotation.
The price points can be compared directly: Individual is $0, Team is $119 per user per month or $99 billed yearly with a 14-day trial, and Enterprise is a quote with VPC or air-gapped installation. Private repositories on the free plan are non-commercial only, the terms were last updated in January 2022, and there is no pipeline orchestration.
What DagsHub does better than Weights & Biases
- Data versioning and annotation on the same platform as tracking
- A published per-user Team price that is easy to model
- VPC or air-gapped install on Enterprise
- MLflow compatibility reduces lock-in on your logging code
Where DagsHub is a step down from Weights & Biases
- Israeli rather than EU contracting entity
- Prices in US dollars per seat
- No pipeline orchestration, and terms dated January 2022
Standout against Weights & Biases. It combines tracking, data versions and annotation, with a per-seat price you can read without a sales call.
Best for: Organisations that must deploy models across hybrid or on-premises clusters
Dutch Analytics B.V. (Chamber of Commerce number 66849381), trading as UbiOps, is based in The Hague. You package a model or pipeline, UbiOps exposes it as an API and schedules it on the compute you attach. The vendor lists API management, version control, scaling, monitoring and auditing, and customer stories include the National Cyber Security Centre of the Netherlands.
It does not replace the tracking part of W&B at all. Teams pair it with DagsHub or ZenML for runs and use UbiOps for what happens after training. UbiOps Cloud is pay per use and UbiOps Private installs in your environment, but no prices are published and the General Terms were last modified in October 2023.
What UbiOps does better than Weights & Biases
- Serving across local, hybrid and multi-cloud infrastructure from one interface
- Private edition installed in your own environment
- Dutch entity with a published register number
Where UbiOps is a step down from Weights & Biases
- No experiment tracking or dataset versioning
- No published prices
- Smaller ecosystem than a hyperscaler serving product
Standout against Weights & Biases. It runs the same deployed model across clusters you operate, which no tracking tool attempts.
Best for: Analyst-heavy organisations that want models built and deployed without code
KNIME AG of Zurich (CHE-114.324.774) is a Swiss company, so it sits outside the EEA although Switzerland has an EU adequacy decision. The Analytics Platform is free and open source, and Business Hub adds scheduling, governance with audit trails, Data Apps and deployment of models or AI agents as services, with plans scaled by capacity instead of per seat.
It is an odd fit for a team leaving W&B, since there is no dedicated experiment tracker or model registry. It suits a different audience: analysts who build in visual workflows and a company that wants governed deployment. Notebook-first teams will find it slower, and Business Hub prices are not published.
What KNIME does better than Weights & Biases
- Free open source desktop with visual workflows
- No per-seat licensing on the Hub
- Audit trails and governance for deployed workflows
Where KNIME is a step down from Weights & Biases
- No dedicated experiment tracking
- Swiss rather than EU entity
- Less natural for notebook-first teams
Standout against Weights & Biases. Analysts can build, schedule and deploy a model in visual workflows with no per-seat licence on the Hub.
Best for: Teams shipping language-model agents or regulated models that need repeatable tests
GISKARD AI SAS, built in France and founded in 2021 according to its About page, tests models and agents rather than tracking them. It scores prompt injection, data disclosure, hallucination and inappropriate refusals, lists 50+ automated adversarial probes on the Professional tier, and connects to CI/CD through an SDK. Teams that used W&B Weave for evaluations will recognise the territory.
The library is Apache-2.0, and the README states that v3 is a rewrite of earlier versions. The enterprise tier lists on-premise, private cloud or SaaS with data residency. Open source support is best-effort, the Professional price is not published, and Giskard does not track training runs or serve anything.
What Giskard does better than Weights & Biases
- Security and quality tests for agents in one tool
- Apache-2.0 library plus on-premise or private cloud deployment
- French entity with data residency options
Where Giskard is a step down from Weights & Biases
- Testing only, no tracking or serving
- Professional price not published
- v3 rewrite may force migration of older scripts
Standout against Weights & Biases. It turns red-teaming into a repeatable test run that CI/CD can gate a release on.
Best for: Computer vision teams that want to cut labeling volume before training
Lightly AG of Zurich (CH-020.3.046.583-0) indexes images with embeddings and offers six sampling strategies to choose which are worth labeling. LightlyStudio adds annotation, QA and an evaluation view with a confusion matrix, and exports COCO, Pascal VOC and KITTI. Teams that used W&B Tables to inspect image data will find the closest equivalent here.
It does nothing for experiment tracking, orchestration or serving, and it handles vision data only. Commercial pricing and data location are not stated on the pages read. Its Swiss entity sits outside the EEA, and LightlyStudio is Apache-2.0 if you prefer to run the code yourself.
What Lightly does better than Weights & Biases
- Embedding-based sampling that reduces what you label
- Apache-2.0 LightlyStudio you can run yourself
- Annotation, QA and evaluation in one tool
Where Lightly is a step down from Weights & Biases
- Vision data only
- No experiment tracking
- Commercial prices and data location not published
Standout against Weights & Biases. It attacks the labeling bill itself, which no tracker does.
Best for: Organisations training across hospitals, banks or devices without sharing raw data
Flower Labs GmbH of Hamburg (HRB 160919) develops an Apache-2.0 federated learning framework and the hosted SuperGrid service. Each participant trains on its own data and only model updates are combined. SuperGrid adds a web UI, public and private federations, audit logs and RBAC, with prices in euros per month billed yearly and 3,000 sign-up credits.
This is not a replacement for a tracker, and it makes sense only if your data truly cannot move. Credit-based pricing is hard to forecast, the company is widening into agents and models, and federations require multi-party agreement that software does not provide. The control plane is hosted by Flower Labs.
What Flower does better than Weights & Biases
- Training without centralising raw data
- Euro pricing on the hosted service
- German entity with a Hamburg register number
Where Flower is a step down from Weights & Biases
- Specialist tool, not a full MLOps platform
- Credit-based pricing is hard to forecast
- Needs coordination between several parties
Standout against Weights & Biases. Raw data stays with each participant, and only model updates travel.
Best for: Teams that already rent GPUs and want hosted MLflow from a Dutch contracting entity
Nebius B.V. of Amsterdam is the contracting entity in its Terms of Use, and the privacy policy also lists subsidiaries in Israel and the United States. Managed Service for MLflow runs standard MLflow as a service, so existing logging code works, and clusters can be created in Finland, Israel, Kansas City, the United Kingdom or France.
For a W&B user who mainly wants a tracking server, this is the least intrusive move. The catch is that it is a compute cloud first: no pipelines, no feature store, prices in dollars, and the MLflow service price was not found on the pages read. An EU-only policy means choosing a Finnish or French region on purpose.
What Nebius AI Cloud does better than Weights & Biases
- Standard MLflow, so little lock-in on logging code
- Region choice including Finland and France
- GPUs and tracking from one European contracting entity
Where Nebius AI Cloud is a step down from Weights & Biases
- Compute cloud first, MLOps second
- US subsidiary listed in the privacy policy
- MLflow service price not published
Standout against Weights & Biases. It is the only tracker here sold by a GPU provider, with a public hourly price list next to it.
What actually breaks when you switch
Start with an inventory, not a trial. List every project that logs to W&B, which of them feed a report someone reads, and which scripts call the wandb library. Decide which projects move, which are archived, and which keep the old tool until they finish. Export metrics, parameters and artifacts for the archived ones before the contract or plan ends, because charts and shared report links do not transfer.
Then expect retraining of habits. Run comparison, tagging and report sharing work differently in each tool, and the dashboards of a MLflow-compatible tracker are plainer than the ones you are used to. Contracts matter too: Valohai, UbiOps, Hopsworks Enterprise and KNIME Business Hub are quoted, so allow time for procurement, and read the current terms, since several public terms pages date from 2021 to 2023.
Can I keep using my existing logging code with a European tracker?
Partly. DagsHub describes its tracking as MLflow compatible, and Nebius runs standard MLflow as a managed service, so code that logs through the MLflow API needs a new tracking address and credentials. Code written against the wandb library calls is different: you either replace those calls with MLflow equivalents or write a thin wrapper that logs to both during the transition.
Valohai and ZenML work differently. Valohai reads metrics printed as JSON and expects inputs and outputs under fixed folders, and ZenML records what steps return. In both cases the logging code you wrote for W&B largely disappears and is replaced by the platform's own conventions.
What happens to my old runs and reports?
Export what you need before the contract ends. Metrics and parameters can usually be pulled through the W&B API into files, and model and dataset artifacts can be downloaded. Charts and shared report links do not travel: they are views inside the W&B interface and have to be recreated in the new tool.
A pragmatic rule is to migrate active projects and archive the rest. Keep a read-only export of finished experiments in cheap storage, move only the runs a current project compares against, and start the new tool with fresh runs for everything else.
Which one to pick
If you mainly need a tracker, choose DagsHub when an Israeli entity is acceptable and you want data versioning next to it, or Nebius AI Cloud when you already rent GPUs and want standard MLflow in a Finnish or French region. If you want to own the pipeline layer under an open licence, choose ZenML, and if your priority is reproducible runs on your own cluster, choose Valohai.
Add the specialists only when the problem is real: Hopsworks for shared features, UbiOps for serving across clusters, Giskard for release testing, Lightly for labeling cost and Flower for data that cannot move. And if you do not handle personal or customer data in your experiments, and the W&B free tier meets your needs, staying where you are is a reasonable decision.
Frequently Asked Questions
For experiment tracking, DagsHub (MLflow compatible, free Individual plan, Team at $99 per user billed yearly) and Nebius AI Cloud (managed MLflow in a European region) are the closest matches. ZenML from Munich and Valohai from Turku cover pipelines and run versioning. Choose by what you track: runs, data, features or deployments.
Weights & Biases Models is sold as part of CoreWeave Forge, and the privacy policy that covers the platform is issued by CoreWeave, Inc., a company based in Livingston, New Jersey. The pricing page at wandb.ai/site/pricing redirects to CoreWeave. We could not read the W&B terms page itself, so we do not state which legal entity signs your contract.
On 23 October 2026 its pricing page listed a free tier, a Pro plan from $60 per month for early-stage teams under 50 employees, and custom Enterprise plans. Check the current page before budgeting, because plan limits and eligibility rules can change.
Yes, within limits. ZenML has a free Apache 2.0 open source plan without an execution cap, DagsHub Individual is $0 with up to 100 tracked experiments in private repositories, KNIME's desktop platform is free, and Hopsworks has a free tier with one project. LightlyStudio and Giskard publish open source code as well.
Yes. ZenML is Apache 2.0 and self-hostable, Valohai deploys on-premises or air-gapped, UbiOps Private installs in your own environment, and Hopsworks publishes a Kubernetes installer for EKS, GKE, AKS and OVHcloud. DagsHub Enterprise supports a VPC or air-gapped install.
Related comparisons
About the author
Explore More European Alternatives
Discover privacy-focused European alternatives to other popular US tech services.