Weights & Biases

Positioning: W&B inside CoreWeave Forge

Weights & Biases needs a status update before a review, because the company no longer stands alone. CoreWeave completed its acquisition of W&B in May 2025, and since 30 September 2026 Free, Pro and Enterprise SaaS accounts sign in through CoreWeave Forge with their existing credentials; self-hosted and dedicated-instance accounts are unchanged. The product set now ships under the CoreWeave Forge banner, which unifies W&B's experiment tracking and evaluation with CoreWeave's Training, Inference, Registry and Sandboxes services and post-training tooling from OpenPipe. CoreWeave's own framing is that accounts, projects, teams and API keys stay the same and W&B stays cloud-agnostic — workloads can live anywhere, not just on CoreWeave iron.

Strip away the corporate history and this remains the experiment-tracking standard for machine learning teams: log training runs, compare them, version datasets and models, ship reports, and — in the Weave line — trace and evaluate LLM applications with the same account. I have used W&B on and off since it displaced TensorBoard in most labs I advise, and this review covers the current combined offering.

What it actually does

The model side is the classic product. Model experiment tracking records metrics, hyperparameters, system stats and artifacts per run, with parallel-coordinates and scatter comparisons across runs, plus Tables for interactive data inspection and Reports for collaborative write-ups with live embedded charts. Artifacts handle dataset and model versioning with lineage tracking across runs. Model registry and lineage features now sit in the platform's asset registry layer.

The GenAI side is Weave: tracing for LLM applications and agents, application evaluations with scorers, PII data redaction, and ingestion metered by data volume. On top of that the Forge-branded additions cover serverless inference for open-source models, Sandboxes for running generated code safely, and ARIA, an AI research and iteration agent. Dedicated scalable deployment is available on Enterprise.

Pricing at time of writing, read from the current pricing page: Free at $0 for personal development, including AI application evaluations, tracing, scorers, model experiment tracking, asset registry and lineage, community support, up to 5 model seats, 5 GB per month of storage, and 1 GB per month of Weave data ingestion. Pro starts at $60 per month billed monthly with a 30-day trial, raising the limits to 10 model seats, 100 GB per month of storage, and 1.5 GB per month of Weave ingestion; additional storage is $0.03/GB and additional Weave ingestion is $0.10/MB. Enterprise is negotiated: configurable seats and storage or bring-your-own bucket, Weave ingestion at $0.20/MB additional, IP allowlisting, private connectivity, CMEK on AWS and GCP, HIPAA possible with BYOB. Forge Inference is billed per model, $5/month on Pro with a limited free credit. There is a free-forever academic licence with Pro features, unlimited tracked hours, 200 GB storage, 25 GB/month Weave ingestion and up to 100 seats. One structural rule catches teams out: Pro is designed for organisations under 50 employees; larger ones are expected to move to Enterprise.

My testing notes

I have instrumented three projects with W&B in the last eighteen months — two training pipelines and one LLM evaluation effort. Four observations.

First, the onboarding experience is still the best in this category. Instrumenting a PyTorch training loop took me about ten lines and a run appeared with system metrics, loss curves and hyperparameter sweeps usable within minutes. When I onboarded a client's team, their first meaningful comparison dashboard existed the same afternoon. That speed is the moat in my experience: the marginal cost of adopting W&B is close to zero, and the switching cost away from it later is not, which is exactly why the free tier is worth taking seriously.

Second, the Weave ingestion meter is the budget line to watch. Weave traces are billed by data volume, and my LLM evaluation project — verbose prompts, retrieval contexts, full agent transcripts — pushed past the Pro inclusion of 1.5 GB/month quickly; overage at $0.10/MB means a heavy tracing month costs real money, not rounding error. I now truncate or hash long retrieval payloads before logging and keep full payloads for sampled runs only. The Free tier's 1 GB/month is a sampler, not a production allowance.

Third, the storage and seat structure is generous at the low end and steep at scale. Free gives 5 GB and 5 model seats; Pro's 100 GB and $0.03/GB additional held fine for my checkpoints because I store large artifacts in my own bucket and reference them — a pattern the platform supports and I recommend. The 50-employee rule on Pro is a real commercial boundary: I have seen a growing client get moved to an Enterprise conversation mid-project, and Enterprise pricing is quote-based.

Fourth, the CoreWeave Forge migration itself was a non-event in my account, which is the best thing I can say about it. Credentials, projects and runs carried over; sign-in moved to CoreWeave Forge. The pricing page now reads "CoreWeave Forge" throughout while the product URLs still live on wandb.ai, and the FAQ still answers in W&B terms, so the brand layer is genuinely mixed today. There is a transitional feel to the surface, but nothing in my workflows broke. Docs still reference W&B product names while the billing pages say Forge, so expect to read both vocabularies for a while.

On comparisons with the obvious alternative, since clients ask: self-hosted MLflow costs nothing in licence and real money in engineering time, and the teams I have seen switch spent their savings on maintenance. W&B's edge is that the experiment tracking, registry and reporting are maintained for you and the integration surface covers the frameworks in daily use. For a team of five shipping models on deadlines, that trade usually resolves in W&B's favour; for a platform team with Kubernetes engineers to spare, MLflow remains a fair choice. I track both in my reviews and pick per team shape rather than on ideology.

What it does not do

This is a tool for people who write training loops and evaluation harnesses; there is no low-code surface for non-engineers. The pricing is tiered and usage-metered in several dimensions at once — seats, storage, Weave ingestion, inference, sandboxes — and the Pro-to-Enterprise step is a sales conversation rather than a self-serve upgrade. Weave's ingestion billing punishes verbose logging unless you curate it. Self-hosted and dedicated instances exist but the story now sits inside CoreWeave's commercial motion; if you chose W&B partly for its independence from any single cloud, the acquisition is a strategic fact to weigh, mitigated by the stated commitment to staying cloud-agnostic. The academic licence is excellent but narrow in eligibility. And older pricing write-ups you will find online describe per-tracked-hour tiers and per-seat rates from the pre-acquisition era — they no longer match the published plans, and I would not budget from them.

Verdict

Weights & Biases remains the experiment-tracking layer I install by default on ML projects, and the Weave line makes the same account cover LLM tracing and evaluation without a second vendor. The Free tier is real, Pro at $60/month is fair for a small team, and the pain starts with uncurated Weave ingestion and the Enterprise step-up. I rate it 4.2: core functionality is deep across tracking, registry, reporting and LLM evaluation, value is good at small scale and meter-dependent at volume, usability is a strength — the fastest instrumentation I know — the ecosystem of framework integrations and community adoption is large, and reliability signals are solid, with the CoreWeave transition as a watch item rather than a problem.

Sources

• https://wandb.ai/site/pricing
• https://www.coreweave.com/companies/weights-and-biases
• https://investors.coreweave.com/news/news-details/2026/CoreWeave-Forge-Launches-Turning-the-AI-Loop-Production-Run-Into-a-Better-Model-and-Agent/default.aspx
• https://www.unite.ai/coreweave-launches-forge-development-layer-for-models-and-agents
• https://www.g2.com/products/weights-biases/pricing
• https://www.spheron.network/blog/weights-biases-pricing-vs-self-hosted-mlflow-2026
• https://www.modern-datatools.com/tools/wandb
• https://fourweekmba.com/ai-coreweave-forge-puts-wb-training-inference-in-one-loop
• https://qa.wandb.ai/site/pricing
• https://www.capterra.com/p/10030581/Weights-Biases