Edge vs cloud inference platforms examines the practical choices teams face when deploying low-latency LLM-powered applications in 2025: run models on-device or at the cloud/edge, or adopt hybrid patterns. The topic centers on on-device LLM inference, local retrieval-augmented generation (RAG), and protocol-driven orchestration (e.g., Model Context Protocol, MCP) that let developers prioritize latency, privacy, connectivity resilience, and cost. Relevance in 2025: hardware acceleration (Apple Silicon, local NPUs), the availability of small but capable foundation models, and growing demand for privacy-preserving, offline functionality have pushed many real-time workloads toward edge and on-prem solutions. At the same time, cloud inference remains attractive for very large models and for workloads that tolerate network round-trips. Key tools and roles: Local RAG — a privacy-first MCP-based document search server that indexes local files (including PDFs) to enable offline semantic search; FoundationModels — an MCP server exposing Apple’s Foundation Models on macOS for local text generation; Minima — an open-source, containerized on-prem RAG stack that can run isolated with embedded LLMs or integrate with services like ChatGPT via MCP; Multi-Model Advisor — an MCP orchestrator that queries multiple Ollama models and synthesizes diverse perspectives for richer answers. These tools illustrate common patterns: local indexing + RAG for context, MCP for client/server interoperability, and multi-model orchestration for robustness. Practical trade-offs: choose on-device for minimal latency and data residency, cloud for scale and model capability, and hybrid architectures to balance the two. MCP-based servers and on-prem RAG stacks are accelerating modular deployments that make these trade-offs explicit and easier to manage.