W&B alternatives: 7 experiment trackers and their real limits

What changed

Weights & Biases is part of CoreWeave now and sold as CoreWeave Forge; wandb.ai redirects there. The wandb Python package is still maintained, at version 0.31.0 as of October 2026.

The Free plan costs $0 and covers up to 5 model seats, 5 GB of storage a month and 1 GB of Weave data ingestion a month. Pro starts at $60 a month with a 30-day free trial, up to 10 seats and 100 GB of storage a month, plus $0.03 for each extra GB. Slack and email alerts are listed as a Pro feature, and academic research stays free with 200 GB of cloud storage.

If you were on Neptune or planning to move there, that option is gone. Neptune agreed to be acquired by OpenAI, turned off its hosted app and API on 5 March 2026, and irreversibly deleted the hosted data that remained.

Compare the trackers

Prices are from each vendor's page, as of October 2026.

The 7 alternatives

MLflow

MLflow 3.17.2, released on 2026-10-21, is open source under Apache-2.0. You run mlflow server yourself on SQLite by default, or point it at PostgreSQL or MySQL, and several clouds sell a managed version.

The real limit is that you run the server and the storage, and there are no alerts for training runs; its webhooks cover registry events. It suits you if you want a free tracker with no seat limits.

Trackio

Trackio 0.44.0, released on 2026-10-27, comes from Hugging Face under the MIT license. It is API compatible with wandb.init, wandb.log and wandb.finish, so import trackio as wandb works for basic logging. Runs live in a local SQLite database, the dashboard can sync to a Hugging Face Space, and trackio.alert() posts to Slack or Discord webhooks.

The limit is stated in the project itself: it's pre-release and may have breaking changes, and anything beyond basic logging isn't covered by the drop-in import. Use it if your scripts already call wandb and you want a local dashboard without rewriting them.

ClearML

ClearML (SDK 2.1.12, Apache-2.0) runs on your own server or on their hosted cloud. The hosted Community tier is $0 for teams up to 3, with 100 GB of artifact storage, 1 GB of metric events and 1M API calls a month. Pro is $15 per user a month plus usage, for teams up to 10.

It fits individuals and small teams who want a hosted UI for less than Forge Pro, or teams ready to run the ClearML server themselves.

Comet

The comet_ml SDK is still actively released, at 3.58.9 on 2026-10-28. Comet's pricing page now centers on Opik, its LLM evaluation and observability product, with Free Cloud at $0 and Pro Cloud at $19 a month.

ML experiment management is listed as an optional add-on with no public price, so you won't know the cost of classic training tracking until you talk to sales. It suits teams already using Opik for LLM work.

TensorBoard

TensorBoard 2.21.0, released on 2026-06-29, is Apache-2.0 and serves on port 6006 on your own machine. The free hosted tensorboard.dev service shut down on 1 January 2024, so sharing curves now means running TensorBoard on a server yourself.

It's still the simplest choice for one person on one machine who wants to watch curves.

DVC with DVCLive

DVC 3.67.1 and DVCLive 3.49.1 are Apache-2.0 tools now maintained under the treeverse GitHub organization. DVCLive writes metrics and plots to files in your git repo, and you compare experiments with dvc exp show in the terminal or in VS Code.

There's no hosted dashboard out of the box. Pick it if you want metrics, plots and parameters versioned next to your code, with no server to run.

Aim

Aim 3.29.1, released on 2025-05-08 under Apache-2.0, gives you a self-hosted UI for comparing runs. There has been no stable release since, so treat it as a maintenance risk and only pick it if you already run it.

How to choose

If you want something self-hosted and free, use MLflow. If you want to keep your existing wandb calls, try Trackio.

If a team wants a hosted UI, compare ClearML with CoreWeave Forge on seat count, since ClearML's free hosted tier stops at 3 users. If you want experiments versioned in git instead of a database, use DVC with DVCLive.

A tracker is not a babysitter

Experiment trackers record metrics so you can look at them later. Most won't tell you a run crashed or hung unless you write alert code yourself, like run.alert() in W&B or trackio.alert() in Trackio. And no alert inside your script can fire when the machine itself dies.

It doesn't draw charts or store metrics, and links expire 30 minutes after the last update, so there is nothing to migrate later. The stall alert fires from the server, so it still arrives when the process is deadlocked or the GPU box is gone.