Weights & BiasesUnclaimed

Weights & Biases is a platform for building AI models and AI applications. Teams use it to track training runs, log metrics and hyperparameters, version datasets and models, and keep a central model registry, which the site describes as the system of record for training AI models.

For AI applications, the Weave product records LLM calls, document retrieval and agent steps. It also provides traces, evaluations, a playground for trying prompts and models, and monitors that run in production.

Developers add the wandb Python library to their code and call wandb.init to start a run. The site shows examples for PyTorch, TensorFlow, Keras, scikit-learn, XGBoost, Hugging Face Transformers, Lightning, LangChain and LlamaIndex. For Weave, the code calls weave.init and marks functions with a decorator so their calls are tracked.

The product is aimed at businesses building AI systems. The site names case studies with Canva, Microsoft and Toyota, and lists industries including autonomous vehicles, financial services, healthcare and life sciences, communications, public sector and scientific research.

The page says the platform is certified under ISO/IEC 27001:2022, ISO/IEC 27017:2015 and ISO/IEC 27018:2019, and is compliant with SOC 2 and HIPAA standards. It says the platform helps customers comply with NIST 800-53 and is in alignment with GDPR requirements. Customers can choose multi-tenant or dedicated deployment.

On the privately hosted pricing page, the Personal plan is listed at $0 per month for personal projects only, with corporate use not allowed. It runs a W&B server locally with Docker and Python. The Advanced Enterprise plan is a custom price and adds single sign-on, audit logs, customer-managed encryption keys, a HIPAA compliant option and an enterprise support package. The page also says academic research is free. Cloud-hosted pricing has moved to CoreWeave Forge.