Loading
Loading
Loading
Prices are our ballpark as of 2026-09-27, not a live store. Memory is what the planner uses.
One DGX Spark holds about a 200 billion parameter model after it is compressed. Two is the size most people finish. Four reaches the largest class and usually needs a fast network switch.
8 of 11 machines
Memory
12 GB VRAM
Price
$180–280 used
What it can run
14B Q4 comfortably. 32B only with a painful offload.
Best for
A used first GPU.
Memory
8 GB VRAM
Price
$250–450 used
What it can run
7B–14B Q4. 32B does not fit.
Best for
A first new card, not an agent box.
Memory
12 GB VRAM
Price
$450–700 used
What it can run
14B easily. A squeezed 32B only if you accept IQ3 and a short context.
Best for
A quiet mid card.
Memory
16 GB VRAM
Price
$800–1,200 used
What it can run
14B with context. 32B Q4 is tight. 70B does not fit.
Best for
A single-GPU coding chat that is not quite a 4090.
Memory
24 GB VRAM
Price
$1,600–2,200
What it can run
Qwen3 32B Q4 with a short context. 70B is a squeeze.
Best for
Private agents: invoices, mail, Slack.
Memory
32 GB VRAM
Price
$2,400–3,400
What it can run
70B-class Q4 with more context than a 4090. Still one model, one box.
Best for
A single-GPU coding model.
Memory
48 GB VRAM combined
Price
$800–1,400 for the pair
What it can run
70B Q4 if the software splits cleanly.
Best for
Tinkerers who like used hardware.
Memory
48 GB VRAM combined
Price
$3,200–4,400
What it can run
70B Q4 with more speed than a 3090 pair.
Best for
A fast private 70B without buying Spark.