Real tests with recorded environments, repeatable commands, measured results and explicit limitations.
Three models, one GTX 1060 3GB and the same 4096 context setting. We measured runtime size, Ollama-reported VRAM residency and generation throughput across 12 runs.
A frozen 2/10 HTML/CSS/JS fixture, one fixed repair task and an unchanged validator. We inspected both returned artifacts and compared the patches.
We changed only the allocated context across six settings and recorded placement, runtime size, aggregate GPU memory and three generation runs at each setting.
What happens when a 4B Q4 model is larger at runtime than the available GPU budget? We measured partial CPU/GPU offload at three context sizes and tested whether freeing Windows VRAM changed placement.
Can a 3 GB Pascal GPU run a current local LLM entirely on the GPU? We installed Ollama on Windows, recorded placement and VRAM snapshots, then ran three API measurements.