Weights & Biases Pricing vs Self-Hosted MLflow (2026)

Weights and Biases pricing runs $60/month for up to 10 seats on its Pro plan, and $315 to $400 per seat per month once you're on Enterprise, according to buyer-reported contracts (Vendr). We've written a lot about the infrastructure underneath training runs, FSDP configs, spot GPU scheduling, checkpoint recovery, but never priced out the tooling layer sitting on top of it. That's a mistake, because for a team running fine-tuning jobs at any real scale, W&B's per-seat line item grows fast enough to matter. This post breaks down what Weights & Biases pricing actually looks like at each tier, what self-hosted MLflow costs on the GPU cluster you already have, and where the breakeven actually sits.

If you're also pricing out the compute side of this equation, our LLM fine-tuning cost breakdown covers API fine-tuning versus renting your own GPUs, and the GPU cost optimization playbook has the broader framework for treating every line item, tooling included, as something you can control.

Weights & Biases Pricing: What It Actually Costs Per Seat at Training Scale

W&B's public pricing covers Free and Pro. Everything past 10 seats or 50 employees moves to a custom Enterprise quote, and that's where the real cost sits: buyer data puts Enterprise at $315-400 per seat per month, with median annual contracts around $47,625 (Vendr). A 10-person team on Pro pays $60/month total, or $720/year, since Pro's price covers the whole 10-seat block rather than scaling per head. The same 10 people on Enterprise, at the reported $315-400/seat/month range, run $37,800-48,000/year, which is the jump that catches most teams off guard once headcount or compliance requirements force the move off Pro.

Free and Pro Tier Limits: The 50-Employee Cutoff and Overage Fees

W&B's Free plan covers up to 5 model seats, 5GB of storage per month, and 1GB of Weave data ingestion per month, positioned for "personal development of AI applications and models" (wandb.ai/site/pricing). It's genuinely free, no credit card, no seat charge, and it's what most solo researchers and small side projects actually run on.

Pro starts at $60/month for up to 10 model seats and 100GB of storage per month, with overage billed at $0.03/GB for storage and $0.10/MB for additional Weave data ingestion beyond the included 1.5GB. W&B's own pricing page is explicit that Pro is "designed for early-stage teams fewer than 50 employees," and once your organization crosses that headcount, you're required to move to Enterprise, independent of how many W&B seats you're actually paying for (wandb.ai/site/pricing).

Cloud marketplace listings add another wrinkle. A single-user Models license runs about $4,800/year on AWS Marketplace, with a separate $25,000 per 12 months for Weave; Google Cloud Marketplace lists a single-user license at roughly $400/month, and Azure pricing mirrors AWS (ZenML). If you're provisioning through a cloud marketplace rather than direct with W&B, run the marketplace number against the direct Pro rate before committing, they aren't identical.

W&B Enterprise Per-Seat Pricing: $315-$400/Seat/Month

W&B doesn't publish an Enterprise rate card. What we have instead is buyer-reported data from Vendr, aggregated across 65 tracked purchases: a median annual contract value of $47,625, a reported range from $11,160/year on the low end to $107,440/year on the high end, and average negotiated savings of 19% off the initial quote (Vendr).

The per-seat number that shows up most consistently in that data is $315-400/seat/month. One buyer's contract is a useful illustration of how W&B's volume tiers actually work: their per-seat rate rose from $315 to $400/unit/month after they cut their license count from 25 seats down to 10 (Vendr). Fewer seats didn't just mean a smaller total bill, it meant a worse rate on the seats that remained. That's the opposite of how most SaaS volume pricing works, and it's worth knowing before you assume trimming a license count is a clean cost lever.

Enterprise buys real capability beyond just "more seats": single sign-on, an optional HIPAA-compliant configuration, single-tenant hosting with a choice of region, customer-managed encryption keys, audit logs, and automated user provisioning, available both cloud-hosted and self-hosted (wandb.ai/site/pricing). For a regulated team, those features can genuinely be worth the premium. For a team that just wants experiment tracking at scale, they're often paying for a compliance package they don't need.

CoreWeave Now Owns W&B: What the $1.7B Acquisition Means for Your Bill

CoreWeave completed its acquisition of Weights & Biases on May 5, 2025, a deal announced that March and reported at roughly $1.7 billion (The Information; CoreWeave). It's the same playbook CoreWeave has run before: buy the software layer that sits closest to your GPU spend, rather than compete purely on chip availability.

Here's the part that matters for your budget specifically: more than a year after the deal closed, W&B still runs on infrastructure from CoreWeave's competitors, with no forced migration and no announced repricing (our earlier coverage of the CoreWeave/Anyscale precedent). That's the encouraging read. The more cautious one: an experiment-tracking platform now owned by a GPU cloud provider is a natural candidate for future bundling, W&B seats packaged into a CoreWeave GPU commitment, priced to make switching away from either harder. Nothing in the public pricing shows that yet, but it's worth watching if your team is negotiating a multi-year contract with either company.

Self-Hosted MLflow on Your GPU Cluster: What You Give Up and Save

Direct answer: MLflow itself has no license fee, it's Apache 2.0 with no paid tiers, so the real cost of self-hosting is the infrastructure you stand up around it: a tracking server, a backend database, and an artifact store. That's typically a small fraction of a single W&B seat per month. What you give up is polish: a sparser UI, no built-in alerting, and no SSO or HIPAA configuration out of the box.

MLflow's Real Infrastructure Cost (Tracking Server, Postgres, Object Storage)

MLflow's architecture has three pluggable pieces: a lightweight FastAPI tracking server that serves the UI and API, a backend store for run metadata (SQLite for a single-host setup, PostgreSQL once you need production-grade concurrency), and an artifact store for large files like checkpoints and model weights (local disk for testing, S3, GCS, or Azure Blob for anything real) (mlflow.org/docs). None of it needs a GPU. It runs fine on a small CPU instance alongside the GPU cluster doing your actual training.

That's a genuinely small bill: a modest VM, a managed Postgres instance, and object storage typically lands well under $100/month, versus $500-$4,000+/month once you're paying for multiple W&B seats. The MLOps overhead of running that stack, patching the tracking server, managing the database, handling backups, isn't zero, and it's the same operational tax our MLOps pipeline orchestration guide covers for self-hosted Kubeflow and Metaflow setups: you trade a subscription line item for engineering time.

If running your own tracking server isn't appealing, managed MLflow hosting exists as a middle ground, Databricks, AWS SageMaker, Azure ML, GCP, and Nebius all offer it (mlflow.org/docs). You still avoid W&B's per-seat pricing, but someone else runs the control plane.

MLflow's adoption numbers back up that it's a mature, well-supported option rather than a niche alternative: over 30 million downloads a month, 20,000+ GitHub stars, and 900+ contributors (mlflow.org). Our own AI infrastructure landscape overview frames MLflow and W&B as the two mainstream options for experiment tracking, and that's still accurate. The difference is what each one costs you to run.

What MLflow Doesn't Do Out of the Box vs W&B (UI, Alerts, Sweeps, SSO/HIPAA)

MLflow's UI is functional but plain, described as "basic compared to commercial alternatives," while W&B's real-time dashboard is routinely called out as the strongest interface in the category (Reintech). If your team lives in dashboards during a training run, watching loss curves update live, that gap is real and not just cosmetic.

MLflow also has no native alerting. You can log metrics, but the rich notification and monitoring layer W&B ships isn't there by default; teams typically script their own alerts for run failures or infrastructure issues on top of MLflow. W&B's Sweeps feature for automated hyperparameter search is another gap: MLflow can log sweep results if you drive the sweep yourself with another tool, but it doesn't have Sweeps' built-in orchestration.

On compliance, self-hosting flips the default in your favor: your experiment data never leaves infrastructure you control, because there's no vendor in the path. W&B's self-hosted option exists too, but it's an Enterprise-only offering, not something available on Free or Pro (Reintech). If data residency is the reason you're evaluating self-hosting in the first place, that's the actual trade: MLflow gets you there on any tier, W&B only gets you there once you're already paying Enterprise rates.

When Weights & Biases Pricing Stops Making Sense for Training Teams

The Breakeven Math: Seats x $/Month vs Self-Hosted Infra Cost

Run the numbers for a 15-person training team on W&B Enterprise at the reported $315-400/seat/month range: that's $4,725-$6,000/month, or $56,700-$72,000/year. The same team's MLflow self-hosted stack, tracking server, Postgres, and object storage, typically runs under $1,200-$3,600/year in raw infrastructure, plus whatever fraction of an engineer's time it takes to keep it patched and available.

The comparison isn't really "$60,000 vs $2,000." It's "$60,000 vs $2,000 plus ops time." If maintaining the self-hosted stack costs less in engineering hours than a single W&B seat's worth of Enterprise pricing, per year, the math tilts self-hosted the moment your team is buying more than one or two seats mostly for logging, not for collaboration or Sweeps. Teams running 10 concurrent fine-tuning jobs at scale already have the GPU-side cost discipline down, our bare-metal H100 fine-tuning case study is a good reference point for what that looks like operationally, and the same discipline applies to the tooling sitting on top of that compute.

A Hybrid Setup: MLflow for Routine Runs, W&B Reserved for High-Value Research

The cleanest version of this isn't all-or-nothing. Run MLflow for the bulk of day-to-day training runs, hyperparameter sweeps you're driving yourself, routine fine-tuning jobs, anything where a researcher just needs to compare loss curves and pick a checkpoint. Reserve a small number of W&B seats for the work where its collaboration features, polished dashboards, and Sweeps orchestration actually earn their keep: cross-team research reviews, leadership-facing reporting, or projects where multiple people need live visibility into the same run.

That split shows up naturally in workloads that already log to both. Our RLHF training infrastructure guide tells readers to log reward and policy metrics to W&B or TensorBoard, exactly the kind of high-stakes, multi-stakeholder run where the richer tooling is worth paying for, even if your routine SFT jobs log to a self-hosted MLflow instance instead. You don't have to pick one platform for the whole org; you have to stop paying Enterprise per-seat rates for logging that doesn't need Enterprise features.

If you're standing up the self-hosted side of that split, Spheron's docs cover provisioning the compute your tracking server and training runs both sit on, and our pricing page has live rates if you're sizing the GPU side of the same budget.

W&B and cloud marketplace figures above are current as of 18 Aug 2026 and are subject to change; verify directly with the vendor before budgeting. Spheron GPU pricing referenced above fluctuates based on availability, check current GPU pricing → for live rates.

Tooling spend adds up the same way GPU spend does, one seat, one instance at a time. If you're already renting GPUs for training, running MLflow's tracking server alongside that cluster costs a fraction of a single W&B seat.

Check current GPU pricing →