Deployment recipe for serving DeepSeek-V4-Flash multimodal inference on two DGX Spark machines, including container images, weights, and networking optimizations for practitioners.