Reproducible experiments, model arithmetic and the limits of both.
Ollama exposes the loaded model, its quantization and its reported VRAM size in one local request.