Weights & Biases

Weights & Biases is the experiment-tracking and AI observability platform used by more than a million machine-learning engineers to log training runs, version datasets and models, evaluate LLM applications and monitor agents in production. Acquired by CoreWeave in May 2025, it now spans classical ML training through production agent evaluation on any cloud.

Weights & Biases is the AI developer platform most large model teams standardised on for experiment tracking, and it now spans the full lifecycle from pre-training to production agents. The Models suite handles run tracking, hyperparameter sweeps, artifact and dataset versioning, a model registry, tables and shareable reports; the Weave suite covers the LLM application layer with traces, evaluations, a prompt playground, monitors and evaluation datasets. Founded in 2017 by Lukas Biewald, Chris Van Pelt and Shawn Lewis, the company reported more than one million AI engineers and 1,400-plus organisations on the platform — including OpenAI, Meta, NVIDIA, Snowflake, AstraZeneca, Toyota, Canva, Square and Wayve — when CoreWeave announced its acquisition on 4 March 2025; the deal closed on 5 May 2025 on undisclosed terms. Since then the two have shipped Mission Control integration, which correlates GPU node failures, hardware errors and networking events directly with training runs, W&B Inference for calling open-weight models such as DeepSeek R1-0528, Llama 4 Scout and Phi 4 Mini through a single endpoint, and Weave Online Evaluations for scoring production agents on any cloud. The open-source Python client on GitHub carries roughly 11,200 stars under an MIT licence and integrates with PyTorch, TensorFlow, JAX and Hugging Face. Deployment is multi-tenant SaaS on Google Cloud, single-tenant Dedicated Cloud on AWS, Google Cloud or Azure, or fully self-managed on any of those plus on-premises bare metal, backed by SOC 2 Type 2, ISO/IEC 27001:2022, 27017 and 27018, HIPAA and GDPR alignment.

Category

Developer Tools

Pricing

Freemium, Subscription, Usage-based, Contact for pricing

Target Market

Data Scientists, ML Engineers, CTOs, Enterprise Developers, AI Platform Teams

Deployment

Cloud-first, Self-hosted, Multi-cloud, Hybrid

Founded

2017

Headquarters

San Francisco, United States

Customers

1,400+ organisations and over 1 million AI engineers (CoreWeave, March 2025)

• ✓
  Experiment tracking and sweeps

  Logs metrics, configuration and system stats for every training run and orchestrates hyperparameter sweeps so results are comparable and reproducible.

• ✓
  Artifact, dataset and model registry

  Versions datasets, checkpoints and models with lineage, so any deployed model traces back to the exact data and code that produced it.

• ✓
  Weave tracing and evaluation

  Captures LLM and agent traces, scores them against evaluation datasets, and provides a playground for comparing prompts and models.

• ✓
  Weave Online Evaluations

  Scores agent performance continuously in production on any cloud, surfacing regressions and feeding new evaluation datasets back into development.

• ✓
  Mission Control integration

  Correlates infrastructure events — node failures, hardware errors, networking issues — directly with the training runs they disrupted, cutting debug time.

• ✓
  W&B Inference

  Serves open-weight models such as DeepSeek R1-0528, Llama 4 Scout and Phi 4 Mini behind one interface, billed per token.

• ✓
  Flexible deployment and residency

  Runs as multi-tenant SaaS, single-tenant Dedicated Cloud in a chosen cloud and region, or fully self-managed on-premises.

• ✓
  Enterprise access governance

  Role-based access control with SSO over OIDC, LDAP and SAML, audit logs, custom roles and customer-managed encryption keys.

• •
  Foundation and fine-tuned model training

  Research teams log thousands of runs, compare configurations and share results as reports rather than screenshots pasted into chat threads.

• •
  LLM application evaluation

  Product teams build evaluation datasets, score prompt and model changes against them, and block releases that regress on quality metrics.

• •
  Production agent observability

  Traces and online evaluations reveal which agent steps fail in production and how often, across whichever cloud the agent runs on.

• •
  Training-run failure diagnosis

  Mission Control ties a stalled or diverging run to the specific node failure or network event behind it, saving expensive GPU hours.

• •
  Model governance and audit

  Regulated teams use registry lineage, audit logs and self-managed or single-tenant deployment to evidence how a deployed model was produced.

List pricing is published only at the small end. Free gives 5 model seats, 5 GB storage and 1 GB/month of Weave ingestion; Pro starts at $60/month for up to 10 model seats, 100 GB storage and 1.5 GB/month ingestion and is explicitly restricted to companies with fewer than 50 employees. The line item that surprises buyers is the overage asymmetry: storage overage is $0.03/GB but Weave data ingestion overage is $0.10/MB — roughly $100/GB — so a trace-heavy LLM application can dwarf the subscription. Everything above Pro, including single-tenant hosting, HIPAA, SSO, audit logs and customer-managed encryption, is custom-quoted Enterprise. W&B Inference and ARIA are billed per token separately, and CoreWeave Sandboxes add $10–$25/month. Self-hosting is free only for a single non-corporate user; academic licences are free with an institutional email.

This page was written from 8 sources, 6 on domains other than wandb.ai.

1. 1.wandb.ai — pricingvendor
2. 2.wandb.ai — securityvendor
3. 3.docs.wandb.ai — hosting
4. 4.github.com — wandb
5. 5.coreweave.com — coreweave to acquire weights biases industry leading ai deve
6. 6.coreweave.com — coreweave and weights biases announce new products and capab
7. 7.investors.coreweave.com — default
8. 8.constellationr.com — coreweave launches latest nvidia instances adds weights bias