Which GPU fits the work you actually do?

Smallest size that is still genuinely useful. Runs on integrated graphics and most laptops without a discrete GPU.

Example fit for an 8GB card. Meta's widely supported 8B instruct checkpoint — Llama 3.2 shipped smaller vision models, not an 8B text upgrade.

Example fit for a 12GB card. Comfortable on a 16GB card with room left for a long context window.

27B open-weight fit example. Google's open family. Does not fit 16GB at Q4 once 20% runtime overhead is included (~16.2GB); plan on a 24GB card.

Mixture-of-experts speed on a single card. Only 3B parameters activate per token, so it runs far faster than its size suggests — but all 30B must still fit in memory.

The practical ceiling for one 24GB card. Needs ~19.2GB at Q4 with 20% overhead — a 24GB card, not 16GB. The last size most single-GPU desktops can hold.

Open multimodal agentic distill for 24–32GB class cards. Catalogued ~29.6B dense with a perception encoder. Plan on official 4-bit / GGUF packs for 24–32GB desktops; full BF16 is not a single-consumer-card fit.

The classic serious-local-setup target. Needs about 35GB of weights at Q4 (~42GB with overhead). Out of reach for every single consumer card; a 64GB Mac is a tight fit with almost nothing spare.

Large MoE that can tight-fit a 96GB workstation card. All 141B parameters must be resident despite 39B active per token — roughly 85GB at Q4 with 20% overhead. That is a tight fit on a 96GB workstation card; 128GB unified memory or a second card is the comfortable path.

Large open-weight reasoning example. 22B active per token keeps it quick, but all 235B must be resident — roughly 141GB at Q4.

Large open-weight coding example. Server hardware only. Around 288GB at Q4 — this is a rack, not a desk.

Permissively licensed generalist example. Around 402GB at Q4. Realistically rented, not owned.