Whether a model runs on your machine is decided first by the graphics card's memory: the model file has to fit in VRAM with room left for the context window and the runtime. The states below are computed from memory alone; software support (CUDA, ROCm, Vulkan) is a separate matter.
Ollama default package 3.4 GB
| VRAM | Local AI | By VRAM |
|---|---|---|
| 32 GB | Comfortable | 1 card |
| 24 GB | Comfortable | 5 cards |
| 20 GB | Comfortable | 1 card |
| 16 GB | Comfortable | 20 cards |
| 12 GB | Comfortable | 15 cards |
| 11 GB | Comfortable | 1 card |
| 10 GB | Comfortable | 2 cards |
| 8 GB | Comfortable | 35 cards |
| 6 GB | Comfortable | 8 cards |
| 4 GB | Runs | 11 cards |
Rule: 1.5× the requirement in memory is Comfortable, 1.15× Runs, 0.6× Limited; the requirement is the default package size, or the maker's stated minimum where there is one (a stated minimum counts as fitting at exactly that figure). Package sizes are the Ollama library's, minimums the maker's own statement; read on 2026-09-11. Sources: ollama.com/library/qwen3.5