This topic covers the techniques, runtimes, and toolchains used to run large language models (LLMs) in memory‑constrained environments—on laptops, edge devices, and privacy‑sensitive local servers. Core approaches include model quantization (reducing numeric precision to 8/4/2 bits or using post‑training/quantization‑aware methods), parameter offloading (moving weights or activations between GPU, CPU and storage), and specialized low‑RAM runtimes and kernels that minimize peak memory and latency. These methods are increasingly relevant as demand grows for local‑first, low‑latency, and cost‑sensitive AI: smaller 3B‑class models (for example, edge‑ready code models) enable on‑device code completion and private assistants without expensive cloud GPU usage. Key tools and categories: LangChain (Agent Frameworks) provides standard model interfaces and orchestration for hybrid pipelines that can combine local and remote inference; Stable Code (Decentralized AI Infrastructure/Research Tools) represents families of smaller, instruction‑tuned code models optimized for fast, private completion; EchoComet, remio, and Znote are examples of local‑first developer and knowledge applications that benefit from privacy‑preserving, memory‑efficient inference by keeping context and model execution on device. Current trends include wider adoption of 4–8‑bit quantization and GPTQ‑style compression, runtime support for NVMe/CPU offloading to trade latency for memory capacity, and integration of these techniques into agent frameworks and data platforms so apps can route workloads between local runtimes and cloud services. Practitioners should weigh trade‑offs among model quality, latency, and operational complexity when choosing quantization and offloading strategies for production workloads.