This topic covers on-device Low‑Rank Adaptation (LoRA) and lightweight LLM frameworks for running and personalizing models locally — comparing adapter‑style approaches such as Tether/BitNet LoRA with alternative on‑device inference and retrieval setups. On‑device LoRA lets users apply small adapter layers to preexisting quantized models to enable personalization and continual learning without full model retraining. Lightweight LLM frameworks focus on reduced memory/compute footprints (quantization, sharded execution, kernel optimizations) so inference and adapters can run on phones, laptops, and edge devices. Relevance in 2026 stems from faster hardware (Apple silicon, energy‑efficient ARM chips), wider adoption of privacy and data‑sovereignty requirements, and the growth of local retrieval‑augmented workflows. Local RAG and Minima demonstrate complementary patterns: privacy‑first semantic search and on‑prem RAG servers that index local documents and surface context to on‑device models. foundation-models (Apple FoundationModels via MCP) shows how OS‑native runtimes integrate into local stacks, while Multi‑Model Advisor highlights multi‑model orchestration for synthesizing perspectives from several local models. The Model Context Protocol (MCP) is emerging as a useful interoperability layer between adapters, local RAG, and native generation backends. Key tradeoffs to evaluate are personalization vs resource use (adapter size, tuning time), model fidelity after quantization, integration complexity with RAG servers, and cross‑platform support. Tether/BitNet LoRA‑style adapters are attractive when you need compact, incremental personalization; alternatives (containerized Minima, MCP-based FoundationModels, Multi‑Model Advisor) are better for integrated on‑prem pipelines, diverse model orchestration, and local document search. Choosing the right approach depends on hardware, privacy requirements, and the need for offline retrieval and synthesis.