CoreWeave Forge Shows 15 Times Faster Model Checkpoint Loading in Test

CoreWeave launched Forge as a software platform connecting AI inference, training, evaluation and agent development, with a preview feature that loads updated model checkpoints into live deployments without a traditional redeployment cycle.

The RL Rollouts feature is available in preview through Dedicated Inference. In a test involving NVIDIA and You.com, hot-loading checkpoints was approximately 15 times faster than traditional redeployment cycles.

The test post-trained NVIDIA Nemotron 3.5 Lightning for web search. BrowseComp accuracy increased from 36.97% to 45.45%, a 22.94% improvement, while average tool calls declined 30.24%.

“Inference becomes part of the training loop itself: the policy generates rollouts, training produces updated weights, and those weights return to serving so the next round of rollouts uses the latest policy,” Urvashi Chowdhary, CoreWeave’s vice president of product and AI services, said.

Inference uses a trained model to produce responses or actions. In reinforcement learning, a model can generate rollouts, receive feedback and undergo repeated updates, making the time required to move new weights into production an important part of the development cycle.

Forge combines Weights & Biases Models, OpenPipe post-training capabilities, the open-source marimo notebook project and CoreWeave services. Its tools include CoreWeave Inference, Agent Lens, Sandboxes, Registry and other services for developing models and agents.

In company benchmarks, the serverless reinforcement-learning service trained models 1.4 times faster and at 40% lower cost than a self-managed setup. Agent Lens detected 20% more critical failures and fixed issues at one-tenth the cost of a general-purpose frontier large language model in another company benchmark.

“CoreWeave Forge brings all these capabilities on one platform, so what a business learns from running AI becomes part of how engineers improve it,” Chen Goldberg, CoreWeave’s executive vice president of product and engineering, said.

Forge launched Sept. 30 during CoreWeave’s Fully Connected 2026 conference in San Francisco. The platform is available in Free, Pro and Enterprise editions, and includes a free 30-day trial of Forge Pro. MasterClass and Canva were already building on Forge at launch.

The launch expands CoreWeave’s infrastructure business into connected software for AI development. CoreWeave’s earlier Forge launch described the platform as covering the development cycle from running and observing workloads to improving and evaluating models.