The rise of open-source generative AI is driving developers to deploy massive models on edge devices like the NVIDIA Jetson platform. This shift enables physical AI agents and autonomous robots to operate independently in the real world. However, a critical challenge remains: running multi-billion-parameter models on systems with limited memory and strict power constraints. Unlike cloud environments, edge devices must share CPU and GPU resources within fixed memory limits, making efficient memory management essential for avoiding latency spikes or system failures. Optimizing memory usage allows developers to enhance performance on existing hardware, support complex workloads like LLMs and sensor fusion, and reduce costs by utilizing smaller memory configurations. NVIDIA has outlined a five-layer optimization framework for Jetson and IGX platforms, potentially reclaiming up to 12 GB of memory while maintaining high accuracy. The foundation lies in the Board Support Package (BSP) and JetPack software stack. By disabling unused services, such as display drivers or camera subsystems when not needed, and adjusting reserved carveout regions at boot, significant DRAM can be freed. Kernel-level optimizations involve tuning the Input/Output Memory Management Unit (IOMMU) and adjusting the Software I/O Translation Lookaside Buffer (SWIOTLB) to eliminate redundant memory reservations. In the user space, developers should identify and terminate background processes that consume unnecessary memory, such as GUI services in headless deployments. Tools like procrank help analyze physical memory usage, allowing teams to reclaim resources from non-essential CPU and GPU allocations. The inference pipeline layer, often managed by frameworks like NVIDIA DeepStream, benefits from disabling visualization stages like Tilers and OSD when only data processing is required. For large language models, efficient serving frameworks like vLLM, SGLang, and Llama.cpp are crucial. These tools utilize continuous batching and KV cache management to maximize throughput. A pivotal strategy is model quantization, which reduces memory footprint by converting weights to lower-precision formats. NVIDIA recommends testing progressively lower precision levels, such as FP8, INT4, or NVFP4, to find the optimal balance between accuracy and efficiency. For instance, using 4-bit quantization on a 2-billion-parameter vision-language model can dramatically lower memory demands. Specialized hardware accelerators further enhance efficiency. The NVIDIA Programmable Vision Accelerator (PVA) offloads always-on vision tasks like motion detection from the main GPU, reducing power consumption and freeing resources for complex inference. A real-world demonstration of these techniques is the Reachy Mini Jetson Assistant. Running on an Orin Nano with only 8 GB of memory, the robot successfully operates a multimodal pipeline concurrently. It utilizes a 4-bit quantized vision-language model via Llama.cpp, speech recognition with Faster-Whisper, and text-to-speech with Kokoro. By combining 4-bit quantization with optimized runtimes and headless operation, the system achieves full functionality without cloud dependency. Overall, these optimizations make it feasible to run LLMs up to 10 billion parameters and vision-language models up to 4 billion parameters on resource-constrained edge hardware. As NVIDIA continues to refine these strategies, developers are increasingly able to bring advanced AI capabilities to the physical world.