Weights & Biases
Experiment tracking and model registry for ML training, plus Weave for tracing and evaluating production LLM agents. CoreWeave-owned since May 2025.
Overview
Two products share one name, which causes some confusion at evaluation time. W&B Models is the experiment tracking system that became the default across ML research: run logging, hyperparameter sweeps, artifact lineage and a model registry, with years of accumulated framework integrations. Weave is the newer LLM and agent side, covering tracing, evaluation, guardrails and a prompt playground. Most teams arrive for one and end up using both.
Weave was rebuilt in 2026 specifically around production agents. Traces are structured as sessions, turns, steps, tools and sub-agents rather than a flat sequence of model calls, which starts to matter the first time you need to work out why an agent looped mid-conversation. Online Evaluations score live traffic and raise alerts through Slack or webhooks when a failure signal moves. There is an MCP server as well, so a coding agent can pull live traces, run an eval and propose a fix without anyone moving data by hand between three tools.
Two things to weigh before signing. Ingestion metering on Weave is aggressive: 1GB a month free, 1.5GB on Pro at $60, then $0.10 per MB, and agent traces are verbose by nature. Teams that instrument everything get a surprise invoice in month two. Second, the company belongs to CoreWeave now, acquired for about $1.7 billion in May 2025. Public commitments to multi-cloud neutrality have held so far. Whether they hold once inference revenue is the incentive is a fair question to put to a sales rep before you standardise on it.
Key Features
- ✓ W&B Models logs training runs, hyperparameter sweeps, artifacts and a versioned model registry, with close to a decade of framework integrations behind it
- ✓ Weave structures traces as sessions, turns, steps, tools and sub-agents rather than a flat list of LLM calls
- ✓ Online Evaluations score live production traffic continuously and fire Slack notifications or webhooks when a failure signal spikes
- ✓ Pre-built scorers for hallucination, PII, toxicity and bias, usable as inline guardrails and not only as offline metrics
- ✓ W&B Inference serves open models including DeepSeek R1 and Llama 4 Scout through one API, running on CoreWeave GPUs
- ✓ An MCP server plus W&B skills let a coding agent read live production traces, run evaluations and iterate without a human shuttling data between tools
Where it holds
- • The training-side product has close to ten years of maturity behind it. Sweeps and artifact lineage still beat what the LLM-native eval startups have managed to build.
- • Weave was rebuilt for agents rather than retrofitted from an LLM logger, so multi-turn session analysis is a first-class view instead of a filter over call logs.
- • Free tier of 5 seats and 5GB storage genuinely covers a small team, and the academic licence is unusually generous at 100 seats.
Where it breaks
- • Weave metering is the trap. 1GB of ingestion a month on free, 1.5GB on Pro, then $0.10 per MB, and verbose agent traces will clear that inside a week.
- • Braintrust is more focused if evals are the whole job, and Langfuse is self-hostable open source, which W&B is not.
- • CoreWeave bought the company for roughly $1.7 billion in May 2025. Neutrality commitments have held so far, but the roadmap now serves a GPU cloud's interests.
- • Two products under one brand means two mental models and some genuinely confusing overlap between MLflow-style run tracking and Weave tracing.
My Take
The 2026 rebuild of Weave is the news here. Traces now model as sessions, turns, steps and sub-agents, which is the correct shape for working out why an agent looped on turn nine of a support conversation and a real improvement over reading a flat call list. Watch the metering closely: 1GB of Weave ingestion a month on free, 1.5GB on Pro at $60, then $0.10 per MB, and a chatty agent will blow through that in days. On the training side, W&B Models is still the thing Comet and MLflow get measured against. The open question is independence, since CoreWeave paid roughly $1.7 billion in May 2025 and a GPU cloud has obvious reasons to care where your inference runs.
Quick Info
- Pricing:
- freemium
- Openness:
- Proprietary
- Starting at:
- Free tier: up to 5 Models seats, 5GB storage and 1GB of Weave data ingestion per month. Pro from $60 a month for teams under 50 people, with 10 seats, 100GB storage at $0.03/GB overage and 1.5GB of Weave ingestion at $0.10/MB overage. Enterprise is custom. Academic licences get Pro features free with unlimited tracked hours, 200GB storage, 25GB/month Weave ingestion and up to 100 seats. W&B Inference is billed per token.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where Weights & Biases holds and where it breaks. The fastest check is your own workload.
Visit website ↗Alternatives to Weights & Biases
Langfuse
open-sourceOpen-source LLM tracing, prompt management and evals. MIT core with no seat or usage caps when self-hosted, cloud from $29 a month. Acquired by ClickHouse in January 2026.
Arize Phoenix
open-sourceSelf-hosted LLM tracing and evals built on OpenTelemetry. Free and uncapped under Elastic License 2.0, source-available rather than OSI open source. Runs in a notebook or on Kubernetes.
Braintrust
freemiumEval-first observability for agents: datasets, scorers and experiments in one loop. Unlimited seats on every tier, Pro at $249 a month flat. SaaS only unless you buy Enterprise.