Weights & Biases is the AI developer platform that ML engineers use to track experiments, tune hyperparameters, version datasets and models, and trace GenAI agents in production. Acquired by CoreWeave in May 2025, it ships as a Python SDK with free, Pro and enterprise tiers plus self-hosted deployment.
Used this tool? Rate it
Used this tool? Rate it
Weights & Biases — almost always written W&B, and reached at the wandb package name that gives this listing its slug — is an AI developer platform whose stated purpose is to be the system of record for AI work. The official site frames the promise as "the AI developer platform to build AI agents, applications, and models with confidence," and the phrase that recurs across its product pages is deliberately archival rather than aspirational: thousands of companies rely on W&B as their system of record for training AI models and developing AI applications.
That framing matters, because it explains what the product is not. It is not a model, not a hosted notebook, and not a drag-and-drop interface for people who do not write code. Every meaningful interaction with W&B begins inside a Python program that is already training something or already calling a language model. Put plainly, the discipline it serves is ML experiment tracking. The tool's job is to observe that program, write down what happened in a durable and comparable form, and make the resulting record searchable months later by a colleague who was not there when the run executed.
The company was founded in 2017 by Lukas Biewald, Chris Van Pelt, and Shawn Lewis, three people who had previously built Figure Eight. The origin story on the company's own About page is unusually specific about the problem they set out to solve: at the time, tracking models was mostly a manual exercise and reproducing them was practically impossible, even as deep learning advanced rapidly around them. What began as an experiment tracking solution developed to help build models at OpenAI grew into an end-to-end MLOps platform. The GitHub repository corroborates the founding date independently — wandb/wandb was created on 24 March 2017, within weeks of the year the company names.
Anyone evaluating W&B in 2026 needs to start with a change of ownership, because it materially affects both the roadmap and the data-processing picture. In May 2025, Weights & Biases was acquired by CoreWeave. That sentence is taken from W&B's own About page, and it is confirmed from the buyer's side: CoreWeave, Inc. (Nasdaq: CRWV) announced from Livingston, New Jersey on 5 May 2025 that it had completed its acquisition of Weights & Biases.
Two things about the transaction deserve careful handling, because the widely circulated version of the story is more confident than the evidence supports.
First, the price. TechCrunch's coverage of the announcement states plainly that CoreWeave acquired the platform for an undisclosed sum, and attributes the frequently quoted $1.7 billion figure to reporting by The Information rather than to either company. The completion press release from CoreWeave, read in full, contains no transaction value at all. So the honest formulation is that the deal value was reported by trade press at around $1.7 billion, against a valuation of $1.25 billion recorded for the company in 2023 — not that the parties disclosed a price. Anyone writing a procurement memo should treat that number as journalism, not as a filing.
Second, the operating entity did not change. The Master Service Agreement still opens by naming Weights and Biases, LLC, a Delaware limited liability company having its principal place of business at 400 Alabama Street, Suite 202, San Francisco, CA 94110. Contracts are still signed with that entity, governed by New York law.
What did change is visible in three places. The leadership page now lists co-founder Lukas Biewald as General Manager and Cofounder rather than chief executive, with Chris Van Pelt and Shawn Lewis carrying Distinguished Engineer titles — an organizational shape consistent with becoming a business unit inside a larger public company. The homepage carries a banner for CoreWeave Sandboxes, isolated environments to run AI agents and model-generated code safely at scale, meaning the parent's infrastructure now appears as a first-class feature inside the W&B surface. And the Serverless Inference product line is described in the vendor's own words as powered by CoreWeave.
For customers worried about lock-in, CoreWeave made an explicit commitment in the completion release: it continues to be committed to the interoperability of the Weights & Biases platform so its customers have the flexibility to choose any deployment option, any infrastructure provider, and any foundation model, framework, or protocol. That is a public promise rather than a contractual term, and it should be read as such — but it is on the record, and it is the most direct answer available to the question of whether buying W&B now means buying into one GPU cloud.
The company describes its team as 270 (and counting) people distributed all over the world. Adoption figures require more care. The number reported by TechCrunch is that over 1,400 organizations, including AstraZeneca and Nvidia, use the tools as their system of record for training and fine-tuning AI models. That is a vendor-supplied figure relayed by media and has not been independently audited; it is quoted here as such rather than presented as verified fact.
The open-source footprint, by contrast, can be checked directly. The wandb/wandb client library carries an MIT license, is written in Python, and at the time of writing shows 11,239 stars, 886 forks, and 954 open issues, with commits landing the same day this page was researched. The separate wandb/weave repository — a distinct product line, discussed below — carries an Apache-2.0 license with 1,121 stars and 226 open issues. Note the license difference between the two repositories: legal review of one does not carry over to the other. Note also the boundary that the star counts can obscure: what is open source is the client SDK and the Weave library, not the hosted service that stores your runs.
W&B is best understood as seven product surfaces sharing one account, one storage quota system, and one Python entry point. The official navigation divides them as Models, Training, Inference, Weave, Registry, Core, and Secure deployment.
Experiment tracking remains the foundation, and the vendor describes it exactly as the archival metaphor suggests: the system of record for your model training, tracking every model, metric, and hyperparameter with just a few lines of code so that you can reproduce results, debug model performance, and optimize faster in a single dashboard.
The instrumentation is genuinely small. A run is opened with wandb.init(project="my_first_project"), hyperparameters are attached to run.config, metrics are pushed with run.log({"loss": ...}), and model files are captured with run.log_artifact(). That last call is where tracking crosses into asset management: an artifact logged during a run is versioned and becomes part of a lineage graph rather than a file sitting on a training node that will be reclaimed tomorrow.
W&B Sweeps handles what most teams call hyperparameter sweeps, and it is one of the rare vendor pages that names its algorithms instead of gesturing at proprietary magic. Bayesian optimization is offered alongside custom logic, and early stopping is implemented with Hyperband specifically to save GPU hours — keeping the most promising, best performing runs alive and killing off the bottom ones so that agents are freed up to try new hyperparameter combinations. The transparency claim is explicit: we cite all the algorithms that we're using, and show all the logs of sweep progress.
For an audience of ML engineers, that disclosure is worth more than any benchmark chart. You can reason about whether Hyperband's aggressive early termination suits your loss curves, which you cannot do with an unnamed "smart search."
Weave is the second product line, and it addresses a different problem than model training. Its entry point is weave.init() plus an @weave.op() decorator applied to a function, after which the platform tracks LLM calls, document retrieval, and agent steps.
The pricing comparison table enumerates what Weave covers: tracing, which monitors and analyzes the performance of GenAI models during development and in production by capturing inputs, outputs, and metadata for each inference; evaluation; production monitoring, which compares different recipes including fine-tuning, RAG, LLMs, and datasets side-by-side for accuracy, latency, and token usage; LLM-as-a-judge metrics; and PII data redaction.
The side-by-side recipe comparison is the feature that most directly reflects how GenAI work actually goes, and it is where LLM observability stops being a buzzword and starts paying for itself. The question is rarely "is this model good" and almost always "is this fine-tune better than that RAG configuration on our data, and at what latency and token cost."
The asset layer carries product names that appear throughout the documentation. Dataset and model versioning is W&B Artifacts. Interactive data visualization is W&B Tables. Collaborative dashboards are W&B Reports. Registry sits above them, acting as the model registry for datasets, models, prompts, code, and metadata with lineage tracking.
Core bundles the remaining infrastructure: Reports, Automations, the SDK, and — a recent addition that signals where the product is heading — Skills and an MCP server for agents.
Inference is the newest surface and the most visibly shaped by the acquisition. Serverless Inference powered by CoreWeave provides API and playground access to leading open-source LLMs, allowing you to develop AI applications and agents without needing to sign up for a hosting provider or deploy models on your own. You can also bring your own trained Low Rank Adaptation (LoRA) weights to run serverless inference with fine-tuned models — which closes a loop that used to require three vendors: train here, register the adapter here, serve it here.
Pricing is per model, with separate rates for input, cached input, and output tokens. Two examples verifiable on the site at the time of writing: NVIDIA Nemotron 3.5 Lightning at $0.10 input, $0.05 cached, and $0.25 output per million tokens with a 262K context window; and DeepSeek V4-Flash-0731 at $0.13 input, $0.07 cached, and $0.28 output, also at 262K context. The model catalogue turns over as new open-weight releases land, so treat the pricing page as the authority rather than any snapshot.
The framework list is broad and each entry ships with working code on the site: LangChain, LlamaIndex, PyTorch, Hugging Face Transformers, PyTorch Lightning, TensorFlow, Keras, scikit-learn, and XGBoost. The Transformers path is the shortest possible: adding report_to="wandb" to TrainingArguments is enough for logging to begin automatically once trainer.train() runs.
The site announces the Weights & Biases mobile app, described as the first iOS app to monitor AI experiments and track training runs anytime, anywhere. This is worth flagging for a reason beyond the feature itself: a request titled "wandb mobile/desktop app?" has been open on GitHub since June 2022, accumulating 27 reactions and 59 comments. It is a case of a long-standing community ask eventually being answered.
The canonical case is a training job measured in GPU-days where the artifact that matters six months later is not the checkpoint but the explanation of how the checkpoint came to exist. Config, metrics, system stats, and the model file all land in one run record. When a regression appears in production, the question "what changed between the model that worked and the model that does not" has a queryable answer.
Sweeps with Hyperband early stopping is a budget instrument as much as an accuracy instrument. Killing the bottom runs early and reallocating agents to fresh combinations is what turns an unaffordable grid search into a tractable one, and the parameter importance visualizations then tell you which knobs were actually load-bearing.
For teams building agents rather than training models, Weave is the entry point. Trace a request end to end through retrieval and model calls, capture inputs and outputs per inference, then compare candidate recipes on accuracy, latency, and token usage before promoting one. PII data redaction exists precisely because production traces of a customer-facing assistant will otherwise contain customer data.
Healthcare, finance, and public sector teams typically cannot send raw training assets to a multi-tenant SaaS. The Bring Your Own Bucket path, discussed in detail under Supported Platforms, exists for this case — with an important carve-out that section spells out.
The free academic license is unusually generous and is discussed under Pricing. For a lab running many small experiments across many students, unlimited tracked hours and 100 seats change the calculus entirely. Read the licensing note in the Limitations section before enrolling sensitive research, however.
Train or fine-tune, version the adapter in Registry, then serve the LoRA weights through Serverless Inference. The consolidation is new and is the clearest product-level dividend of the CoreWeave combination so far.
1. Install and authenticate. pip install wandb, then log in with your API key. Note one known friction point documented by the community: an issue titled "wandb.login() hangs and then disconnects me every time I try on Google Colab" has been open since April 2021 with 15 reactions and 38 comments, so if you work primarily in hosted notebooks, budget a little time for the auth handshake.
2. Open a run and record your configuration. Call wandb.init(project=...) at the top of your training script and assign hyperparameters to run.config. Do this before training rather than after — the config is what makes runs comparable later, and a run logged without it is a chart without an explanation.
3. Log metrics inside the training loop. run.log({"loss": loss}) on whatever cadence suits your job. The site's PyTorch example logs every args.log_interval batches, which is the right instinct: log often enough to see shape, rarely enough not to drown the dashboard.
4. Capture artifacts. run.log_artifact("./my_model.pt", type="model") puts the model into the versioned lineage graph. If your weights already live in cloud storage, use reference artifacts — data stored in an external bucket does not count against your storage quota, only the metadata does.
5. If you use a supported framework, take the shortcut. With Hugging Face Transformers, report_to="wandb" in TrainingArguments is the whole integration. With Lightning, instantiate WandbLogger and pass it to the Trainer. Hand-rolled logging is unnecessary in most stacks.
6. Move to Sweeps once single runs are boring. Define the search space and let the agents work. Pick your own distributions for inputs and enable early stopping so the GPU budget goes to the promising region.
7. Add Weave for anything with a model call in it. weave.init("your-project") plus @weave.op on the functions you care about. The decorator approach means you are annotating existing code rather than restructuring it.
8. Set up Reports and Automations for the humans. A Report is where a training result becomes something a colleague reads. Automations and Slack or email alerts (available on Pro and above) are what let you stop watching the dashboard.
9. Decide on deployment before you scale, not after. Multi-tenant, dedicated, or customer-managed — and if data residency is a constraint, read the BYOB section carefully, because its coverage is not uniform across product lines.
ML engineers and research scientists are the primary audience and the reason every entry point is a Python import. If your work is a training script, this is instrumentation for that script.
GenAI application developers arrive through Weave rather than through experiment tracking. They may never log a single training run and still get full value from tracing and evaluation.
ML platform and infrastructure teams care about the deployment matrix, BYOB scopes, SSO, SCIM provisioning, custom roles, and audit logs — the parts of the product that make it defensible to a security review.
Academic researchers get a free Pro license with terms detailed below, subject to the data-usage caveat in Limitations.
Individual practitioners and learners can work indefinitely on the free tier, or run a local server with wandb server start — though the local Personal tier carries a licensing restriction that matters (see Limitations).
Who it is not for: anyone looking for a no-code AI tool. There is no meaningful path through this product that does not involve writing Python. Teams that need data residency for GenAI traces specifically should also read the BYOB carve-out before committing.
W&B runs wherever your training code runs — it is a library plus a service, not an application with system requirements. What varies is where the service itself lives.
The pricing page lists SaaS (multi-tenant cloud), Dedicated, and Customer-managed side by side. Enterprise adds a single-tenant option with choice of region.
The bar is deliberately low: to self-host W&B, you need Docker and Python installed, and the server runs on Linux, macOS, or Windows systems that meet those requirements. Startup is two commands — pip install wandb then wandb server start.
The responsibility trade is stated plainly by the vendor: self-hosting gives you full control over your data and privacy, but you are responsible for managing your data storage and backup. That is the honest version of the deal and worth repeating to anyone who assumes self-hosting is strictly safer.
BYOB lets you store artifacts and other sensitive data in your own cloud or on-premises infrastructure. For Dedicated Cloud and Multi-tenant Cloud, W&B does not copy the data you store in your bucket into W&B-managed infrastructure. Supported back ends are CoreWeave AI Object Storage, Amazon S3, Google Cloud Storage, Azure Blob Storage, and S3-compatible storage such as MinIO Enterprise (AIStor) hosted in your own environment. Communication between the SDK, CLI, UI and your buckets uses pre-signed URLs.
Two caveats decide whether this satisfies a real compliance requirement, and both come from the documentation rather than the marketing pages.
The first is scope by deployment. Multi-tenant Cloud supports team-level BYOB only; instance-level is not applicable there. Dedicated Cloud and Self-Managed support both instance and team level.
The second is more consequential: the W&B Platform secure storage connector, or BYOB, is not available for Weave. If your data residency requirement covers GenAI application traces — the inputs and outputs of production LLM calls — BYOB does not currently address it. This is the single most important limitation for regulated teams evaluating the GenAI side of the platform, and it is easy to miss because the pricing table renders it as a terse footnote reading "Not available with Weave."
Even with BYOB enabled, data is split. The central database retains metadata for users, teams, artifacts, experiments and projects, plus Reports, experiment logs, system metrics, and console logs. Your bucket holds experiment files and metrics, artifact files, media files, run files, and exported history metrics and system events in Parquet format. "Bring your own bucket" does not mean "nothing leaves your perimeter."
An iOS app exists for monitoring experiments and training runs. Site localization is partial: link enumeration of the marketing site reveals Japanese, Korean, and German sub-sites only. There is no Chinese, French, Russian, Spanish, or Portuguese localized site, so non-English-speaking teams outside those three languages should expect to work in English documentation.
Two separate tracks exist — cloud-hosted and privately-hosted — and each has its own free tier with different terms. Confusing them is easy and consequential.
Free — $0/mo. Designed for personal development of AI applications and models. Includes AI application evaluations, tracing, scorers, AI model experiment tracking, AI assets registry and lineage tracking, and community support. Quotas: up to 5 model seats, 5 GB/mo storage, 1 GB/mo Weave data ingestion.
Pro — starts at $60/month, billed monthly, with a 30-day free trial. Adds unlimited teams for collaboration, team-based access controls, service accounts, priority email and chat support, CI/CD automations, and Slack and email alerts. Quotas: up to 10 model seats, 100 GB/mo storage with additional storage at $0.03/GB, and 1.5 GB/mo Weave data ingestion with additional ingestion at $0.10/MB.
Enterprise — custom plans. Adds a single tenant option with choice of region, a HIPAA compliant option, secure private connectivity, customer-managed encryption keys, single sign-on, automated user provisioning, custom roles, audit logs, and an enterprise support package.
An important qualifier on Pro that is a licensing rule rather than a technical limit: W&B Pro is designed for early-stage teams and organizations with fewer than 50 employees, and customers who exceed these guidelines will be required to transition to W&B Enterprise, because Pro is designed with rate limits and performance constraints that may not support higher usage. A 60-person company cannot simply keep buying Pro seats.
Personal — $0/mo. One user seat, experiment tracking, registry and lineage tracking, running locally on any machine with Docker and Python. The restriction is unambiguous and repeated twice on the page: for personal projects only, corporate use is not allowed.
Advanced Enterprise — custom plan, with a free enterprise trial license available, adding flexible deployment options, HIPAA compliance, secure private connectivity, CMEK, SSO, automated user provisioning, custom roles, audit logs, and an enterprise support package.
Free forever for academic research: unlimited projects, teams, and 200 GB of free cloud storage. Eligibility is a free Pro license for academic institutions pursuing research not connected to a for-profit entity, intended for students, professors, and postdoctoral researchers, and requiring an active email address affiliated with an academic institution. The license includes all Pro features, unlimited tracked hours, 200 GB of cloud storage, up to 25 GB/mo of Weave data ingestion, and up to 100 seats, with extra storage at $0.03 per GB billed monthly.
Pro is billed monthly or annually upfront. Model seats are prorated if added mid-term, though no credits are issued for removed seats. Weave data ingestion, W&B Inference, and storage are billed monthly in arrears based on usage. Enterprise plans are invoiced annually upfront.
Storage is computed as a 30-day weighted average using the GB-days arithmetic given above, and covers both artifacts and data logged to runs. Reference artifacts and data in external buckets do not count. Inference is billed per model, with free credits offered for a limited time and Pro including a $5/mo credit, also time-limited. CoreWeave Sandboxes is in public preview with $10/mo and $25/mo credits on Free and Pro respectively, likewise for a limited time — treat all four of those credit lines as promotional and verify current terms before budgeting.
Free gets community support. Pro gets priority email and chat with a first response of 4 to 24 hours depending on severity. Enterprise gets 1 to 4 hours depending on severity and support package. Business hours are defined as 2am to 5pm PT, Monday to Friday. A dedicated Slack or MS Teams channel and a dedicated AI solution engineer are included only with Premium Support.
Choosing an experiment tracker is mostly a question of what else you are willing to adopt alongside it.
MLflow is the open-source default and the most common comparison. If your requirement is that the tracking server be entirely yours with no vendor relationship, MLflow answers it. W&B's counter-argument is the hosted collaboration layer, the Sweeps implementation, and the GenAI tracing side, none of which MLflow matches out of the box.
Neptune.ai and Comet occupy adjacent positions as commercial trackers. Evaluation between them tends to hinge on team size pricing and specific visualization needs rather than on capability gaps.
TensorBoard remains the zero-friction local option. It is the right answer for a single person on a single machine, and it stops being the right answer the moment a second person needs to see the results.
Cloud-native platforms — Vertex AI, SageMaker, Azure ML — bundle tracking into the platform where the training already runs. The trade-off is portability, and it is exactly the trade-off CoreWeave publicly promised not to impose on W&B customers.
LangSmith, Langfuse, Arize Phoenix compete with Weave specifically rather than with W&B as a whole. Teams that only build LLM applications and never train models should evaluate that narrower field on its own terms.
A structural note on comparison: because the W&B client is MIT-licensed and Weave is Apache-2.0, the instrumentation code you write is not a proprietary dead end even though the hosted backend is. That reduces, though it does not eliminate, the cost of changing your mind later.
This section is longer than most because the constraints here are documented rather than speculative, and several of them live in the Master Service Agreement rather than on the pricing page.
Free and academic customer data may be used for testing and development. The MSA states that W&B may use Customer Data from free and academic customers for testing and development purposes. This is the most important sentence on this page for anyone considering the free tier for unpublished research or commercially sensitive work, and it appears nowhere on the pricing page. Paid tiers are not covered by this clause. Read it before enrolling a lab's pre-publication experiments.
The privately-hosted Personal tier forbids corporate use. For personal projects only. Corporate use is not allowed. An engineer running a local server on a work laptop for a work project is outside that license.
Free plans carry no continuity commitment. W&B reserves the right to modify, discontinue, suspend or terminate any free subscription plans at any time in its sole discretion, without prior notice.
BYOB does not cover Weave. Repeated here because it is the limitation most likely to derail a late-stage compliance review. GenAI trace data cannot currently be kept in your own bucket.
AI features and outputs come with no IP indemnity. The agreement is explicit: W&B makes no representation or warranty that AI Features or Output will be free from third-party intellectual property and is not obligated to indemnify the customer against infringement claims related to them. All AI features and output are provided "as is." Separately, the customer acknowledges that due to the nature of machine learning, outputs may be inaccurate or identical across users, and the customer will evaluate all output before relying on it.
Beta features carry no warranty, indemnity, SLA, or support, and may be discontinued at any time without liability. Since Sandboxes is currently in public preview, this clause is live and not hypothetical.
Runs are effectively append-only. The community evidence is consistent and long-running: renaming groups of completed runs (open since 2020, 66 reactions), merging old runs with new (2022, 51 reactions), copying a run between projects (2023, 37 reactions), and reusing a deleted run name (2020, 26 reactions) are all still open requests. Plan your project structure accordingly.
Notebook authentication can be fragile. The Colab login issue has been open since 2021 with 38 comments.
Customer names and logos may be used in marketing. The agreement has the customer consent to W&B's use of their name and logos in marketing materials, public announcements, and investor communications.
Data export is your responsibility and timing matters. The customer is solely responsible for exporting Customer Data before expiration or termination, after which W&B deletes it per its policies. Payment obligations are non-cancellable and fees non-refundable except as expressly provided.
Broad warranty disclaimer. Outside the specific warranty section, the W&B assets are provided "as is" with performance, merchantability, fitness for purpose and non-infringement expressly disclaimed, and no warranty that the service is error-free or uninterrupted.
Governing law is New York, with exclusive venue in New York County and a jury trial waiver. Non-US customers should factor dispute-resolution cost into procurement.
Not for children under 13.
No verifiable third-party rating exists. This is a limitation of the available evidence rather than of the product, and it is stated plainly rather than papered over. The Trustpilot profile for wandb.ai is unclaimed with a 0.0 score and zero reviews — the page literally reads that the company has not received any reviews yet. Attempts to verify ratings on G2, Gartner Peer Insights, SourceForge, and Slashdot were all blocked at fetch time, so no rating from those sources is quoted anywhere on this page. The absence is unsurprising for developer infrastructure whose users express themselves through GitHub issues rather than review sites, but it means the third-party evidence here rests on media reporting and verifiable repository data instead of scores. As a related caution: the unclaimed Trustpilot profile lists a company location of Anguilla, which contradicts the San Francisco address in the operating agreement; that listing should not be relied on.
Data processing is US-centric. The subprocessor table lists AWS, Azure, GCP, and Clickhouse as processing in the United States or as chosen by Customer, while CoreWeave (listed as an Affiliate, for cloud infrastructure administration), Datadog, and Hex are listed as United States. Note that the parent company now appears in the processing chain in a position that customer region selection does not change.
The bug bounty scope is narrow. Only findings on qa.wandb.ai and api.qa.wandb.ai qualify; all other subdomains are excluded. Awards run from $50–$100 for low severity to $750–$1,000 for critical issues such as SQL injection, RCE, privilege escalation, or SSRF to an internal service. Researchers must sign an NDA and may not publicly discuss findings without written approval — a policy that limits the public security signal available to prospective customers.
For a platform whose buyers are frequently security teams rather than engineers, the certification list is specific and worth reproducing accurately. The platform is certified under ISO/IEC 27001:2022, ISO/IEC 27017:2015, and ISO/IEC 27018:2019, and remains compliant with SOC 2 Type 2 and HIPAA standards. It also helps customers comply with NIST 800-53 and is in alignment with GDPR requirements for processing personal information.
Encryption is TLS 1.2+ in transit and AES 256 at rest. Authentication supports single sign-on via OIDC, LDAP, or SAML, with role-based access control for granularity. Security testing combines regular vulnerability testing with penetration testing, and the compliance program includes both internal reviews and independent third-party assessments.
Feature availability by tier is uneven and should be checked against your requirements: SOC 2 Type II and customizable data retention via Artifacts TTL appear across tiers, while IP allowlisting, secure private connectivity (inbound and outbound), and customer-managed encryption keys for AWS and GCP are marked as available with private hosting. HIPAA compliance is marked as possible with BYOB. Custom roles and bring-your-own-bucket are both marked as not available with Weave.
The legal documentation set is published at a single entry point and includes Website Terms of Use, the Master Service Agreement, W&B AI Terms, a Service Level Agreement, a Support Policy, a Data Processing Addendum, the Privacy Policy, and a Modern Slavery Statement. The existence of a standalone DPA and SLA as public documents is itself a procurement-relevant maturity signal.
CoreWeave. The acquisition completed on 5 May 2025, announced by CoreWeave, Inc. (Nasdaq: CRWV), and confirmed on W&B's own About page. The contracting entity is unchanged — agreements are still with Weights and Biases, LLC, a Delaware LLC based in San Francisco. Co-founder Lukas Biewald is now listed as General Manager and Cofounder.
No, and CoreWeave said so publicly in the completion release: it remains committed to the interoperability of the platform so customers can choose any deployment option, any infrastructure provider, and any foundation model, framework, or protocol. Treat that as a public commitment rather than a contract term, and note the practical counterweight: CoreWeave now appears in the subprocessor table as an affiliate handling cloud infrastructure administration.
The cloud Free plan is $0/mo with AI application evaluations, tracing, scorers, model experiment tracking, registry and lineage tracking, and community support. Ceilings are up to 5 model seats, 5 GB/mo storage, and 1 GB/mo of Weave ingestion. It stops at team collaboration — unlimited teams, team-based access controls, service accounts, CI/CD automations, and alerts all require Pro. The most important non-quota limit is contractual: W&B may use Customer Data from free and academic customers for testing and development purposes.
Yes, and the terms are specific. It is a free Pro license for academic institutions pursuing research not connected to a for-profit entity, for students, professors, and postdoctoral researchers, requiring an active institutional email. It includes all Pro features, unlimited tracked hours, 200 GB of cloud storage, up to 25 GB/mo of Weave ingestion, and up to 100 seats. Extra storage is $0.03 per GB monthly. The same testing-and-development data clause that applies to free accounts applies here.
They are separate product lines with separate entry points and separate repositories. wandb handles model training — experiment tracking, sweeps, artifacts, registry — and its client is MIT-licensed. weave handles GenAI applications — tracing LLM calls, evaluation, production monitoring, LLM-as-a-judge metrics, PII redaction — and is Apache-2.0. They share an account and billing, and Weave has its own ingestion quota measured separately from storage.
Partially. Bring Your Own Bucket supports CoreWeave AI Object Storage, Amazon S3, Google Cloud Storage, Azure Blob Storage, and S3-compatible stores like MinIO Enterprise. But BYOB is not available for Weave, so GenAI traces cannot be kept in your bucket. And even where it applies, the central database still holds metadata for users, teams, artifacts, experiments and projects, plus Reports, experiment logs, system metrics, and console logs.
Storage is a 30-day weighted average, not a peak. The vendor's own worked example: 100 GB for 15 days, 200 GB for one day, and 300 GB for 14 days gives 5,900 GB-days, divided by 30 equals 196.6 GB, rounded to 197 GB for the month. The main lever is reference artifacts — data held in an external bucket does not count against the quota, only the metadata stored in W&B does.
Yes. You need Docker and Python, and the server runs on Linux, macOS, or Windows. Startup is pip install wandb then wandb server start. What you take on is stated by the vendor: you get full control over your data and privacy, but you are responsible for managing your data storage and backup. Also note the licensing split — the free Personal tier is for personal projects only and corporate use is not allowed, so business self-hosting requires the Advanced Enterprise plan.
Because none could be verified first-hand. The Trustpilot profile for wandb.ai is unclaimed, scores 0.0, and has zero reviews. G2, Gartner Peer Insights, SourceForge, and Slashdot were all blocked during research, so no score from them is quoted. Rather than repeat an unverified number, this page relies on what can be checked: repository metrics, open community issues, published pricing, and the contract text.
The client libraries are; the hosted service is not. wandb/wandb is MIT-licensed with 11,239 stars, 886 forks, and 954 open issues, created in March 2017. wandb/weave is Apache-2.0 with 1,121 stars and 226 open issues. Both were receiving commits on the day this page was researched. You can read and fork the instrumentation code, but the SaaS backend that stores and serves your runs remains proprietary.
