Weights & Biases (Weave)

One-liner — The ML developer platform for experiment tracking and model management, whose Weave product does LLM/agent observability and evaluation; acquired by GPU-cloud CoreWeave in 2025.

Categories — llm-observability

What it does

W&B started as the standard tool for ML experiment tracking (logging runs, metrics, hyperparameters, datasets, model versions) and grew into a model/AI developer platform. Its LLM-relevant product is Weave: tracing, logging, evaluation and monitoring of LLM and agent applications — capturing prompts, responses, tool calls, latency/cost, and running evals. The job here: see what your LLM apps actually did, measure quality/regressions, and debug agent behavior.

Where it sits in the stack

Model/prompt layer, llm-observability. It is observability/evaluation, not an inline guardrail; but evals and tracing are how you catch unsafe or degraded model behavior before and after release. Adjacent to ai-red-teaming (eval overlap) and complementary to inline ai-runtime-security.

Deployment & architecture

SaaS, plus dedicated/self-managed and on-prem options; primary integration is via SDK instrumentation in training/inference code. Framework-agnostic; works across clouds and model providers. CoreWeave has committed to keeping the platform interoperable (any cloud/infra/model/framework) rather than locking it to CoreWeave GPUs.

Positioning & differentiators

Known for being the default experiment-tracking tool in the ML research community (OpenAI among customers) and for Weave’s eval-centric LLM observability. Nearest neighbors in LLM observability: langfuse, langsmith, arize-phoenix, braintrust, helicone, datadog, comet (Opik). W&B’s edge is the breadth from classic ML training through LLM/agent eval; the open question is neutrality now that it is owned by a GPU cloud.

Ownership, funding & M&A

Verified. Acquired by CoreWeave (Nasdaq: CRWV) — announced 2025-03-04, closed 2025-05-05. Deal value not officially disclosed; widely reported at ~$1.7B (The Information). Founded 2017 by Lukas Biewald, Chris Van Pelt, Shawn Lewis; HQ San Francisco; raised ~$250-300M pre-deal (investors incl. Insight Partners, Felicis, Coatue). Now a CoreWeave subsidiary; CoreWeave states it will keep the platform interoperable. Ownership confidence high on acquirer/dates; price reported, not confirmed.

CTO / hedge-fund lens

Day-1 if you build or fine-tune models / ship LLM apps in-house; not needed if you only consume third-party AI assistants. For a quant/ML-heavy fund, W&B/Weave is a strong choice for experiment tracking + LLM eval and supports self-hosting for data-control reasons. The CoreWeave ownership is the main diligence point: assess data-residency, neutrality, and whether you are comfortable with a GPU-cloud vendor owning your ML metadata plane. Relevant to model-risk/validation evidence under SR 11-7-style processes (eval and lineage records).

Competitors / alternatives

langfuse, langsmith, arize-phoenix, braintrust, helicone, datadog, comet, fiddler-ai.

Open questions / to verify

• Official deal price (only reported, not disclosed).
• Concrete data-residency/neutrality guarantees post-CoreWeave for non-CoreWeave customers.

Sources

• CoreWeave Completes Acquisition of Weights & Biases — fetched 2026-06-28 — supports: acquirer CoreWeave, close 2025-05-05, interoperability commitment; confidence: high
• CoreWeave acquires AI developer platform Weights & Biases (TechCrunch) — fetched 2026-06-28 — supports: announce 2025-03-04, ~$1.7B reported, OpenAI customer; confidence: med
• About Weights & Biases / Crunchbase — fetched 2026-06-28 — supports: founded 2017, founders, HQ, funding; confidence: med

History

• [2026-06-28] Stub created from seed registry.
• [2026-06-28] Researched; verified CoreWeave acquisition (announced 2025-03-04, closed 2025-05-05, ~$1.7B reported), raised ownership_confidence low→high; filled founding/HQ/funding, clarified Weave = LLM observability product.