CoreWeave has introduced CoreWeave Forge, a new platform designed to integrate model serving, observability, post-training, and evaluation tools. The service is available for free to start, with additional paid tiers offering expanded capabilities for developers.
The company also unveiled CoreWeave RL Rollouts, a feature built on Nvidia Corp.’s Dynamo framework. This tool is designed to accelerate the iteration of agentic models by allowing new checkpoints to be loaded into live deployments. In internal testing, the solution reduced model reload latency by 15x.
The provider is shifting its focus from raw GPU capacity toward a full-stack approach, including managed services for training and inference. This strategy addresses the growing demand for inference, which, according to a company survey, is projected to reach a 50% workload share for some healthcare clients within 12 months.
CoreWeave’s optimization strategy involves tuning layers above the hardware, such as the vLLM engine, quantized models, and custom speculative decoders. The company emphasizes the use of open-source technologies to ensure customer flexibility.
