Wandb
Wandb: The AI developer platform. Use Weights & Biases to train and fine-tune models, and manage models from experimentation to production.
Last verified:
What is Wandb?
Wandb is the leading AI developer platform for training and fine-tuning models, managing models from experimentation to production, and tracking and evaluating GenAI applications powered by LLMs. It serves as a comprehensive system of record for machine learning teams, covering the entire ML lifecycle from experiment tracking to model deployment.
Wandb pricing
Pricing model: Freemium
Free tier: $0/mo includes up to 5 model seats, 5 GB/month storage, 1 GB/month Weave data ingestion, AI application evaluations, tracing, scorers, experiment tracking, registry & lineage tracking, and community support. Pro plan: Starts at $60/user/month billed monthly (30-day free trial available), includes up to 10 seats, 100 GB/month storage, 1.5 GB/month Weave data ingestion, unlimited teams, team-based access controls, service accounts, priority email & chat support, CI/CD automations, Slack/email alerts. Additional storage $0.03/GB, additional Weave data $0.10/MB. Enterprise: Custom pricing with customizable seats, storage, and data ingestion, single tenant option, HIPAA compliance, SSO, SCIM, audit logs, custom roles, customer-managed encryption keys, enterprise support package. Academic license: Free Pro features with 200GB storage, unlimited tracked hours, up to 100 seats for accredited institutions. Personal plan: $0/mo for 1 user, personal projects only, no corporate use allowed.
Wandb pros
- Experiment tracking with real-time metrics visualization
- Interactive dashboards for comparing multiple runs side-by-side
- Model registry with versioning and lineage tracking
- W&B Weave for agent AI tracing and evaluation
- Hyperparameter sweeps for automated tuning
- Artifact management for datasets and model checkpoints
- OpenAI-compatible Inference API for foundation models
- Serverless RL training for fine-tuning LLMs
- Collaborative reports for team sharing
- SDK integration with TensorFlow, PyTorch, Keras, scikit-learn
- LangChain and Llama Index support for GenAI
- CI/CD automations for ML workflows
- Mobile app for monitoring experiments on iOS
- SOC 2 Type II, HIPAA, ISO 27001 compliance certifications
- Free tier with 5 GB storage and 5 seats
- Community support and extensive documentation
- PII data redaction for GenAI applications
- LLM-as-a-judge metrics for cost and performance monitoring
- Offline mode with later sync capability
- Git integration automatically logs commit information
Wandb cons
- No easy way to delete individual runs from UI
- Storage overages charged at $0.03/GB can add up
- Weave data ingestion overages at $0.10/MB are expensive
- Pro plan limited to teams under 50 employees
- Tracked hours billed at $1/hour beyond 5,000 annual limit
- Free tier limited to 5 GB monthly storage
- Free tier Weave data ingestion only 1 GB/month
- Inference credits are limited time promotional offer
- Self-hosting requires Docker and Python setup complexity
- Rate limits on Pro plan may not support high usage
Frequently asked questions about Wandb
What is Weights & Biases used for?
Weights & Biases is an AI developer platform used for experiment tracking, model management, and GenAI application evaluation. It helps ML engineers track hyperparameters, metrics, and artifacts during training; manage models from experimentation to production via the registry; evaluate and monitor LLM applications with Weave; access foundation models through Inference API; and fine-tune models with serverless RL Training. It serves as a central system of record for the entire ML lifecycle.
How much does Weights & Biases cost?
W&B offers a Free tier at $0/mo with 5 seats and 5 GB storage. The Pro plan starts at $60/user/month billed monthly with 10 seats and 100 GB storage. Enterprise plans have custom pricing based on team needs. Additional storage costs $0.03/GB and additional Weave data ingestion costs $0.10/MB. Academic Research license is free with 200GB storage and up to 100 seats.
What is W&B Weave?
W&B Weave is a product for developing GenAI applications with tracing, evaluation, and monitoring capabilities. It captures inputs, outputs, and metadata for each inference, allows comparison of different recipes (fine-tuning, RAG, LLMs, datasets) side-by-side for accuracy, latency, and token usage, provides LLM-as-a-judge metrics, and includes PII data redaction. It integrates with LangChain, Llama Index, and OpenAI.
Does Weights & Biases have a free tier?
Yes, W&B has a Free tier designed for personal development and small projects. It includes AI application evaluations, tracing, scorers, experiment tracking, registry & lineage tracking, up to 5 model seats, 5 GB/month storage, 1 GB/month Weave data ingestion, and community support. Tracked hours are limited by the storage limit.
What frameworks does W&B support?
W&B supports major ML frameworks including TensorFlow, PyTorch, Keras, scikit-learn, XGBoost, and Lightning. For GenAI, it integrates with LangChain, Llama Index, and OpenAI. The SDK works with any Python code, and there are specific callbacks and loggers for Keras (WandbMetricsLogger, WandbModelCheckpoint), PyTorch Lightning (WandbLogger), and TensorFlow.
Is Weights & Biases compliant with security standards?
Yes, W&B is certified under ISO/IEC 27001:2022, ISO/IEC 27017:2015, and ISO/IEC 27018:2019, compliant with SOC 2 Type II and HIPAA standards, helps customers comply with NIST 800-53, and aligns with GDPR requirements. Enterprise plans offer additional security features like single sign-on, customer-managed encryption keys, audit logs, and secure private connectivity.
How is storage calculated in W&B?
Storage includes both artifacts and data logged to runs. W&B calculates usage over the last 30 days using GB-days: multiply storage amount by days used, sum all GB-days, then divide by 30. For example, 100 GB for 15 days plus 300 GB for 14 days equals (100×15 + 300×14)/30 = 196.6 GB, rounded to 197 GB. Only metadata stored in W&B counts, not references to external storage.
What is the difference between Free and Pro plans?
Free includes up to 5 seats, 5 GB storage, 1 GB Weave data, and community support. Pro includes up to 10 seats, 100 GB storage, 1.5 GB Weave data, unlimited teams for collaboration, team-based access controls, service accounts, and priority email & chat support. Pro also includes all Free features like CI/CD automations, Slack/email alerts, and AI application evaluations. Pro is designed for teams under 50 employees.
Who qualifies for the academic license?
The free academic license is for students, professors, and postdoctoral researchers at accredited educational institutions pursuing research not connected to for-profit entities. An active email address affiliated with an academic institution is required. It cannot be used for commercial purposes, government research, or consulting for commercial, governmental, or nonprofit organizations.