NVIDIA has introduced ModelExpress, a new framework designed to accelerate the distribution and loading of large language model weights across distributed GPU clusters. The system addresses a persistent infrastructure bottleneck: the recurring latency and network overhead incurred when moving terabyte-scale model checkpoints between remote storage, local disks, and GPU memory. ModelExpress optimizes the weight lifecycle through intelligent path selection and peer-to-peer data transfer. For initial cluster ingress, the framework streams checkpoints directly from object storage or local disks into GPU memory using NVIDIA ModelStreamer and GPUDirect Storage, eliminating intermediate host-memory staging and redundant disk writes. Once a primary replica completes its load, ModelExpress automatically shifts to direct peer-to-peer transfers. Leveraging the NVIDIA Inference Xfer Library, the system enables GPU-to-GPU communication over RDMA. Compatible replicas discover each other through metadata backends like Redis or Kubernetes, exchange transfer credentials, and pull weights directly from already-initialized GPU memory. This fan-out architecture ensures each new worker joins an expanding source pool, eliminating repeated cold starts and network ingress bottlenecks. Beyond raw weight distribution, ModelExpress mitigates inference warmup latency by sharing compiled kernel caches and runtime artifacts. When a new replica joins a deployment, it receives pre-compiled inference kernels and optimized library artifacts directly from a serving peer via CPU-to-CPU RDMA. This approach removes the requirement for shared storage volumes and allows the entire fleet to bypass the multiminute JIT compilation and autotuning phases that traditionally dominate startup times. The framework also extends this receiver-driven architecture to reinforcement learning post-training workflows. As trainers continuously update policy weights, rollout workers automatically discover and fetch the latest model partitions and metadata, converting distributed training layouts into inference-ready tensors without interrupting generation cycles. In benchmarks conducted on eight NVIDIA B200 GPU nodes equipped with ConnectX-7 networking, ModelExpress reduced the cold start time for the DeepSeek-V4 Pro model from eight minutes to one minute and forty-four seconds. Artifact and weight transfers between replicas were completed in under ten seconds. The system maintains hardware and software agnosticism by automatically probing available capabilities and falling back to host-staged I/O or direct storage reads when specific transfer paths are unavailable. The framework is already integrated with major inference engines including vLLM and SGLang, and maintains native compatibility with open-source initiatives like Dynamo and llm-d. By automating weight distribution and abstracting hardware-specific transfer backends, ModelExpress establishes a standardized pathway for scaling large-model deployments, significantly lowering infrastructure costs and accelerating model iteration cycles in production AI environments.