Weights & Biases

Overview

Weights & Biases, usually written W&B, is a platform for tracking and sharing machine learning work. Its best-known product records every training run's settings, metrics, system statistics and outputs, and turns them into live dashboards that a whole team can explore. Around that core it offers hyperparameter sweeps, versioned datasets and models, a registry, written reports, and a separate toolkit called Weave for tracing and evaluating LLM applications.

An everyday picture: imagine a running club where each member wears a fitness watch. Every run's pace, heart rate and route is uploaded automatically, and the coach can put any two runs side by side, spot who is improving and share a summary with the club. W&B does that for model training. Each run reports in as it happens, and anyone on the team can compare, comment and publish what they found.

W&B is mainly offered as a hosted service, with dedicated cloud and self-managed options for organisations that need them. Since 2025 it has been part of CoreWeave, a GPU cloud provider.

What it is

W&B is organised into two main product families, plus supporting services.

The parts share one data model. A run consumes artefacts and produces new ones, so W&B can draw a lineage graph from raw data to deployed model.

Where it came from

Lukas Biewald, Chris Van Pelt and Shawn Lewis founded Weights & Biases in 2017 in San Francisco. Biewald and Van Pelt had earlier founded a data-labelling company, and the new venture grew from the observation that deep learning practitioners were tracking experiments in spreadsheets and notebooks.

The first product was experiment tracking with a lightweight Python library. Sweeps, artefacts, reports and a model registry followed as teams used it for larger projects, including training of large language models. In 2023 and 2024 the company added Weave, aimed at developers building applications on top of LLMs rather than training models.

In March 2025 CoreWeave announced it would acquire Weights & Biases, and the deal completed in May 2025. W&B continues as a product line within CoreWeave.

How it works

Instrumenting a run

A training script initialises a run, records its configuration and logs metrics as training progresses.

import wandb

run = wandb.init(project="churn", config={"lr": 0.001, "epochs": 10})
for epoch in range(run.config.epochs):
    loss = 1.0 / (epoch + 1)  # replace with real training
    run.log({"epoch": epoch, "loss": loss})
run.finish()

Integrations with frameworks such as PyTorch, PyTorch Lightning, Keras and Hugging Face Transformers log much of this automatically. The library also records GPU and CPU usage, the git commit and the command line.

Comparing in dashboards

Runs stream to the W&B server, where a project workspace shows every run as lines on charts. Users group runs by any config value, filter, add panels and build parallel-coordinates plots to see which settings matter.

Sweeps

A sweep is defined by a search space and a method: grid, random or Bayesian. A central controller hands out configurations to agents running on any machines, which run training and report back. Early-termination rules can stop weak runs.

Artefacts and registry

Datasets and models are logged as artefacts with automatic versioning and content hashing. A run that uses dataset:v3 and produces model:v12 is recorded in the lineage. Chosen versions are linked into the registry with aliases such as production, ready for deployment systems to fetch.

Weave for LLM applications

Weave wraps functions in an application, such as retrieval, tool calls and model calls, so each request is captured as a trace. Teams build evaluation datasets, score outputs with code or an LLM judge, compare versions of prompts and models, and monitor production traffic.

Key concepts and terms

A worked example

This scenario is illustrative. A team is fine-tuning a vision model to classify defects on circuit boards.

1. They log the labelled image set as artefact boards:v1 with 18,000 images.
2. A Bayesian sweep explores learning rate, augmentation strength and batch size across 40 runs on four GPU machines, each running a sweep agent.
3. The sweep dashboard shows that augmentation strength has the largest effect, and that learning rates above 0.003 diverge. Early termination stops 14 weak runs, saving about a third of the compute.
4. The best run reaches 96.8 per cent validation accuracy. Its model is logged as defect-model:v9, and the lineage graph shows it was trained on boards:v1 with commit 7e2d1f0.
5. Reviewers write a report embedding the confusion matrix and a sample of misclassified images, then link the model into the registry with the staging alias.
6. Two months later, new images are added as boards:v2. A retrain on the new data is compared directly with the earlier run in one chart.

Strengths and limitations

Strengths:

• Very quick to adopt: a few lines of code give rich dashboards.
• Excellent visual comparison and collaboration features.
• Sweeps distribute easily across machines.
• Artefact lineage connects data, runs and models.
• Weave extends the same habits to LLM application development.

Limitations:

• The core service is commercial; free tiers have limits, and large teams pay per seat or usage.
• Hosted by default, which some regulated organisations cannot use without a dedicated or self-managed deployment.
• Logging very high-frequency metrics from many runs can slow dashboards.
• It does not orchestrate pipelines or serve models itself.
• Being part of a cloud provider may matter to organisations that want strict vendor neutrality.

How it compares with alternatives

Practical guidance and best practices

1. Put every hyperparameter in config, so runs can be grouped and filtered later.
2. Log datasets as artefacts before training, so lineage is complete from the start.
3. Use clear project and run naming, plus tags for experiment purpose.
4. Log metrics at a sensible frequency; every step is rarely needed for long runs.
5. Use sweeps with early termination for any search larger than a handful of runs.
6. Write a report for each decision, so the reasoning survives beyond the dashboard.
7. Use registry aliases for deployment rather than hard-coded artefact versions.
8. Keep API keys in environment variables or secret stores, never in code.

Common mistakes

• Logging metrics without the config, which makes comparisons meaningless.
• Leaving offline runs unsynced, so results never reach the team.
• Storing huge datasets as artefacts when references to object storage would do.
• Running sweeps over too many dimensions at once with too few runs.
• Using latest as the deployment alias, so any new upload goes to production.
• Treating dashboards as documentation instead of writing down conclusions.

Frequently asked questions

Is Weights & Biases free?

There is a free tier for individuals and academic use, with paid plans for teams and enterprises. Pricing and limits change, so check the current plans before committing.

Can I self-host W&B?

Yes. Besides the multi-tenant cloud, W&B offers dedicated cloud and self-managed deployments for organisations with data residency or security requirements.

What is the difference between W&B Models and Weave?

Models covers training work: experiments, sweeps, artefacts and the registry. Weave covers building applications on LLMs: tracing requests, evaluating outputs and monitoring production behaviour.

Does W&B work with any framework?

The Python library works with any code that can call wandb.log. There are ready integrations for popular frameworks, which reduce the logging code needed.

How is W&B different from MLflow?

Both track experiments and manage models. W&B is a commercial, mainly hosted service known for visualisation and collaboration. MLflow is open source, easier to self-host and governed by a foundation. Many teams choose based on hosting needs and budget.

Glossary

• Workspace: the dashboard view of a project's runs.
• Panel: a single chart or table in a workspace or report.
• Parallel coordinates: a chart showing how several settings relate to a metric across runs.
• Lineage: the graph linking artefacts to the runs that created and used them.
• Early termination: stopping weak sweep runs before they finish.
• Monitor: a Weave feature that scores production traffic continuously.
• Offline mode: logging locally and syncing to the server later.

Summary

Weights & Biases records training runs as they happen and turns them into shared, interactive dashboards, with sweeps, versioned artefacts, a registry and reports built on top, and Weave extending the approach to LLM applications. Founded in 2017 and part of CoreWeave since 2025, it is popular for its speed of adoption and collaboration features. Teams should weigh its commercial, mainly hosted model against open-source options, but for many it is the quickest route to seeing and sharing what their models are doing.

Related

It belongs to MLOps in my stack, under MLOps and AI infrastructure. Others in the same group: