CoreWeave launches Forge for AI model and agent development
Forge brings training, agent observability, evaluation and inference into one platform, with several components available immediately.
Forge connects the AI development loop
CoreWeave launched Forge on September 30 as a development platform for teams building and improving models and agents. Its announcement says the platform brings training, experiment tracking, agent traces, evaluations and inference into a connected workflow. It combines Weights & Biases Models, OpenPipe post-training capabilities, marimo-based notebooks and CoreWeave’s cloud services. CoreWeave says teams can still use models, frameworks and clouds beyond its own, and names MasterClass and Canva as existing Forge customers. Those customer and interoperability statements are the company’s claims in its launch announcement.
What is available
CoreWeave says Forge is available in Free, Pro and Enterprise editions. It lists ARIA, a coding agent for analyzing experiment and observability data, and isolated Sandboxes as generally available. Agent Lens adds observability for production agents; RL Rollouts, a feature for updating checkpoints during inference, is in preview. The company also describes versioning for model checkpoints and agent configurations through CoreWeave Registry.
The launch establishes product availability, not independently measured gains. CoreWeave publishes its own comparisons for Agent Lens and post-training speed and cost, but the reviewed source does not provide outside validation. Whether the integrated workflow reduces handoffs for teams using multiple clouds remains an open practical question.