Trusted Partners
Nearly everything people run locally rests on one of three foundations. The llama.cpp engine and its ggml library sit underneath Ollama and LM Studio, read models in the GGUF format, and will use a graphics card, the processor cores, or a split between them. PyTorch and vLLM read safetensors weights, expect a supported card, and are what most published research code and serving stacks assume. ONNX Runtime is the third, and it is the only one of them that can reach a neural processing unit.
That last point is worth stating plainly, because it is the most common misunderstanding about AI-branded machines. Installing Ollama or LM Studio on a laptop sold on the strength of its neural processing unit will not touch that unit; the work goes to the graphics portion of the chip or to the CPU cores. Using the NPU means converting a model to ONNX and compiling it for the vendor's execution provider, Qualcomm's QNN or Intel's OpenVINO, which is a deliberate project rather than a setting you toggle.