TL;DR — Key Takeaways
- CoreWeave launched Forge, a new development layer combining Weights & Biases Models, technology from OpenPipe and marimo, and CoreWeave’s own AI infrastructure services.
- Forge adds new and expanded capabilities including Agent Lens in public preview, generally available ARIA and Sandboxes, and RL Rollouts for Dedicated Inference.
- Weights & Biases remains cloud-agnostic, while CoreWeave’s managed compute services give the company a path to bring more AI development workloads onto its own infrastructure.
CoreWeave has launched CoreWeave Forge, a new development layer that pulls model training, inference, evaluation, experiment tracking and agent development into a single environment, bringing together technology from several companies the AI cloud provider has acquired over the past two years.
Forge combines Weights & Biases Models with technology from reinforcement learning startup OpenPipe and the open source marimo notebook project, along with CoreWeave’s own inference, training, registry and sandbox services. The platform is available now in Free, Pro and Enterprise editions.
When CoreWeave acquired Weights & Biases in May 2025 for roughly $1 billion, the deal gave it an established software platform used throughout the AI development lifecycle. CoreWeave began integrating the two businesses within weeks of closing, including making infrastructure health data from its Mission Control service available within W&B Models, where it could be correlated with individual training runs. Forge now brings that into a single product layer. Existing Weights & Biases SaaS customers are also being moved into the Forge account and navigation environment, while retaining their existing credentials, projects and API keys.
CoreWeave describes Forge as a way to connect what it calls the AI improvement loop: run a model or agent in production, observe its behavior, turn production signals into datasets and evaluations, improve the system, evaluate the new version and deploy it again. The company says these stages are often handled across separate tools and workflows, and Forge can connect them and carry relevant data into each round of model or agent development.
Several capabilities are launching or changing with Forge. CoreWeave Agent Lens is a new public preview service that analyzes production agent traces to identify failures and turn them into evaluation cases. ARIA, CoreWeave’s AI research and iteration agent, is now generally available, along with CoreWeave Sandboxes. Dedicated Inference is also adding RL Rollouts in preview, so a training loop can keep running without redeploying, the company said.
Nick Patience, vice president and practice lead for AI platforms at The Futurum Group, said teams choosing AI platforms have typically had to decide how much flexibility they are willing to give up in exchange for a more connected toolchain. He said Forge is built around avoiding that tradeoff by connecting production back into training within one environment while staying open across models, frameworks and clouds. Doing both at once is a more difficult engineering challenge than solving either side separately, he added.
Forge is not entirely infrastructure neutral, however. Weights & Biases was already cloud-agnostic before the acquisition and remains that way under Forge, but managed services including CoreWeave Training, Inference and Sandboxes run on CoreWeave compute.
Strategy-wise, that gives CoreWeave a way into AI development workflows even when the underlying workloads are running somewhere else. Weights & Biases can still serve as the software layer across teams’ existing infrastructure, while Forge gives CoreWeave, whose revenue still primarily comes from cloud computing services, an opportunity to move some of the more compute-heavy work onto its own services and infrastructure.
