Weights & Biases
MLOps and LLM observability, now inside CoreWeave Forge
- Overview
- Weights & Biases (W&B) is an MLOps platform acquired by CoreWeave in May 2025 and now part of the CoreWeave Forge development layer. It covers experiment tracking, artifact versioning, hyperparameter sweeps, and LLM observability via W&B Weave — and remains cloud-agnostic despite the acquisition. New capabilities added post-acquisition include Serverless SFT, isolated agent Sandboxes, and the ARIA research agent.
- Why it matters
- W&B's acquisition by CoreWeave (closed May 2025) is the clearest sign that GPU hyperscalers are moving up the stack to own the developer tooling layer, not just compute. Teams already on W&B get access to tightly integrated Sandboxes and serverless fine-tuning on CoreWeave infrastructure, but the platform remains cloud-agnostic — a deliberate choice to protect its broad installed base. The pricing restructure (Pro now $60/mo with usage-based metering) means costs can scale non-linearly as Weave ingestion and Inference usage grow; buyers should model this before committing. The main competitive risk is that hyperscaler-native alternatives (Microsoft Foundry, Google's Gemini Enterprise Agent Platform) now bundle similar tracking capabilities, squeezing W&B's standalone value proposition.
- In practice
- ML teams typically instrument training runs with a few lines of the wandb Python SDK, then use W&B Weave to trace and evaluate LLM pipelines in the same workspace — avoiding a separate observability tool. Post-acquisition, teams running RL or agent eval workloads can spin up CoreWeave Sandboxes directly from the W&B UI without managing separate infrastructure. The main honest limitation is that deep Sandbox and Inference features are tied to CoreWeave compute, so teams on other clouds get a narrower feature set than the marketing implies.
- Best for
- ML engineers and research teams who need reproducible experiment tracking, LLM observability, and fine-tuning workflows on any cloud infrastructure
- Pricing
- Free tier (5 seats, 5 GB storage); Pro $60/mo (10 seats, 100 GB, usage metering on extra storage, Weave ingestion, and per-token Inference); Enterprise custom pricing
Key features
- Experiment tracking with automatic logging and baseline comparisons
- W&B Weave: LLM tracing, evaluation, and dynamic leaderboards
- Artifact and model registry with lineage tracking
- Sweeps for hyperparameter optimization
- CoreWeave Sandboxes for isolated agent and RL workload execution
Quick answers
- What is Weights & Biases?
- Weights & Biases (W&B) is an MLOps platform acquired by CoreWeave in May 2025 and now part of the CoreWeave Forge development layer. It covers experiment tracking, artifact versioning, hyperparameter sweeps, and LLM observability via W&B Weave — and remains cloud-agnostic despite the acquisition. New capabilities added post-acquisition include Serverless SFT, isolated agent Sandboxes, and the ARIA research agent.
- Why does Weights & Biases matter?
- W&B's acquisition by CoreWeave (closed May 2025) is the clearest sign that GPU hyperscalers are moving up the stack to own the developer tooling layer, not just compute. Teams already on W&B get access to tightly integrated Sandboxes and serverless fine-tuning on CoreWeave infrastructure, but the platform remains cloud-agnostic — a deliberate choice to protect its broad installed base. The pricing restructure (Pro now $60/mo with usage-based metering) means costs can scale non-linearly as Weave ingestion and Inference usage grow; buyers should model this before committing. The main competitive risk is that hyperscaler-native alternatives (Microsoft Foundry, Google's Gemini Enterprise Agent Platform) now bundle similar tracking capabilities, squeezing W&B's standalone value proposition.
- What is Weights & Biases best for?
- ML engineers and research teams who need reproducible experiment tracking, LLM observability, and fine-tuning workflows on any cloud infrastructure
- How much does Weights & Biases cost?
- Free tier (5 seats, 5 GB storage); Pro $60/mo (10 seats, 100 GB, usage metering on extra storage, Weave ingestion, and per-token Inference); Enterprise custom pricing
Seen in recent stories
Where Weights & Biases showed up in the last 90 days of KeyNews editions.
Know the terms
Glossary entries that explain what Weights & Biases is built on.
Evals
Systematic evaluation frameworks that measure AI model performance on specific tasks relevant to your use case, going beyond generic benchmarks to test the behaviors that actually matter for your application.
Fine-tuning
The process of continuing to train a pre-trained model on a smaller, task-specific dataset. Fine-tuning customizes model behavior for specific domains or formats and is a key part of most enterprise AI deployments.
Benchmark
A standardized test used to compare AI model performance. Common benchmarks include MMLU, HumanEval, and GSM8K. While useful for ranking, benchmarks can be gamed and may not reflect real-world value.
AgentOps
The operational discipline of deploying, monitoring, debugging, and managing AI agents in production. AgentOps encompasses observability, cost tracking, failure recovery, human-in-the-loop escalation, and compliance auditing for autonomous AI systems.
More enterprise ai tools
New tools drop every week.
ONE BRIEFING · EVERY FRIDAY · FREE
Free. Unsubscribe anytime.