Model fit
Which machine runs which model
Memory decides which models fit. Pick a model size and quality, and see every machine and card we sell that can hold it.
30–34B at 4-bit
About 20 GB of GPU memory
The sweet spot for many people: noticeably smarter answers, and still a single 24–32 GB card at 4-bit.
Examples: Gemma 2 27B, Qwen2.5 32B
13 of 15 machines and cards can run it
Intel Arc Pro B6024 GB$599.99
AMD Radeon RX 7900 XTX24 GB$990
AMD Radeon AI PRO R9700 (32 GB)32 GB$1,299
Tenstorrent Wormhole n30024 GB$1,399
Tenstorrent Blackhole p150a32 GB$1,399
NVIDIA GeForce RTX 409024 GB$1,761
NVIDIA GeForce RTX 509032 GB$1,999
Minisforum MS-S1 Max128 GB unified$2,719 – $3,523
AMD Radeon PRO W7900 (48 GB)48 GB$3,999
Apple Mac Studio (M3 Ultra)96 GB unified$4,404 – $10,682
NVIDIA DGX Spark128 GB unified$4,699 – $10,593
Beelink GTR9 Pro128 GB unified$4,790
NVIDIA RTX PRO 6000 Blackwell (96 GB)96 GB$14,594
All numbers are approximate. We count about (parameters × bits ÷ 8) for the weights plus 20–30% for context and buffers, then treat 90% of a graphics card's memory and 75% of a unified-memory system's memory as usable. Long contexts, image models and batching need more. One card is counted on its own; several cards can be combined in one machine.