What Is Forge in the CoreWeave Documentation?

Updated

Forge is the section of the CoreWeave documentation that covers developing and evaluating AI workloads, rather than running the underlying infrastructure. It is the place to look if you want to track experiments, evaluate and improve agents and LLM applications, prototype in notebooks, or manage models across their lifecycle. If your question is about provisioning GPUs, clusters, storage, or networking, that lives under Infrastructure instead.

What Forge covers

The documentation groups Forge around a single purpose: "Develop and evaluate." Within that, it lists four broad capabilities:

  • Track experiments — record runs, parameters, and results.
  • Evaluate and improve agents and LLM applications — measure quality and iterate.
  • Prototype in notebooks — explore data and model workflows interactively.
  • Manage models across their lifecycle — version, store, and hand off models and datasets.

The tools inside Forge

Tool What it does
Weights & Biases (W&B Models) Experiment tracking, hyperparameter sweeps, artifacts, reports, and automations.
W&B Weave LLM and agent evaluation, alongside W&B Models.
ARIA AI research agent that analyzes experiments, explains results, and builds visualizations and reports. Marked as Preview.
Agent Lens Intelligent agent observability and improvement.
Notebooks Reactive Python notebooks for exploring data, prototyping model workflows, and analyzing runs and artifacts.
Registry Versioned hub for models and datasets with lineage and deployment handoff.

The relationship between these is layered rather than parallel. W&B Models handles the general experiment-tracking layer, and Weave sits alongside it specifically for LLM and agent evaluation. ARIA and Agent Lens build on top of that recorded data — ARIA to interpret experiments and produce visualizations and reports, Agent Lens to observe and improve agents. Notebooks and Registry are the working surfaces: you prototype in notebooks, and you version and hand off what you produce through the registry.

How Serverless services fit in

Forge also lists Serverless services, described as a way to "serve, fine-tune, and run models on CoreWeave infrastructure without managing it." Three are named:

  • Post-Training — post-train and fine-tune LLMs with Serverless RL and SFT, and distill specialized models from real application traffic.
  • Serverless Inference — open-source foundation models through an OpenAI-compatible API, integrated with Weave.
  • Serverless Sandbox — on-demand, isolated compute environments you create and discard from Python.

These sit in Forge because they are part of the model development and evaluation loop rather than the infrastructure layer. Serverless Inference being integrated with Weave is the clearest link: you can serve a model and evaluate it in the same workflow.

Where to start

If you are new to Forge, the order implied by the documentation is:

  1. Notebooks — prototype and explore data or model workflows.
  2. W&B Models and Weave — track experiments and evaluate LLM or agent behavior.
  3. ARIA and Agent Lens — once you have runs recorded, use these to analyze results and improve agents. Note that ARIA is labeled Preview.
  4. Registry — version what you want to keep and hand off for deployment.
  5. Serverless services — serve, fine-tune, or run the model without managing infrastructure.

If you instead need to set up the environment these run on — instances, capacity, regions, billing, or policies — start from the Infrastructure section of the docs. Forge assumes that layer is already in place.