One number decides it: usable memory for the model. Pick a model and a quality level — every device below is checked against that single feature first, then compared on everything else: speed, screen, weight, ports. Prices live on the detail pages and at the retailer; the ranking never depends on them.
Usable memory is what is left for the model itself. A graphics card uses all of its VRAM; a unified-memory device shares memory with the system, so about 75 percent of it can hold a model. The bar shows the selected model against that usable memory.
96 GB usable. Holds a 70B model at Q4 quality — unified memory, no spill to slow system RAM.
36 GB usable. Holds a 32B model at Q4 quality — unified memory, no spill to slow system RAM.
24 GB usable. Holds a 32B model at Q4 quality — unified memory, no spill to slow system RAM.
24 GB usable. Holds a 32B model at Q4 quality — dedicated vram memory, no spill to slow system RAM.
24 GB usable. Holds a 32B model at Q4 quality — dedicated vram memory, no spill to slow system RAM.
16 GB usable. Holds a 14B model at Q4 quality — dedicated vram memory, no spill to slow system RAM.
12 GB usable. Holds a 14B model at Q4 quality — dedicated vram memory, no spill to slow system RAM.
8 GB usable. Holds a 8B model at Q4 quality — dedicated vram memory, no spill to slow system RAM.
Pick up to three. The winner of each row is marked; the answer underneath says why, starting with the memory feature and then the rest of the device.
Sorted by usable model memory — the feature that decides which models a device can run at all.
| # | Device | Usable | Largest at Q4 | Price |
|---|---|---|---|---|
| 1 | MacBook Pro 16in — M4 Max, 128GB | 96 GB | 70B | — |
| 2 | MacBook Pro 14in — M4 Pro, 48GB | 36 GB | 32B | — |
| 3 | MacBook Pro 14in — M4, 32GB | 24 GB | 32B | — |
| 4 | Razer Blade 16 (2025) — RTX 5090 | 24 GB | 32B | — |
| 5 | Lenovo Legion Pro 7i — RTX 5090 | 24 GB | 32B | — |
| 6 | ASUS ROG Zephyrus G16 — RTX 5080 | 16 GB | 14B | — |
| 7 | ASUS ROG Zephyrus G14 — RTX 5070 Ti | 12 GB | 14B | — |
| 8 | RTX 5070 laptop (budget class, many brands) | 8 GB | 8B | — |
Usable memory is what is left for the model itself. A graphics card uses all of its VRAM; a unified-memory device shares memory with the system, so about 75 percent of it can hold a model. The bar shows the selected model against that usable memory.
32 GB usable. Holds a 32B model at Q4 quality — dedicated vram memory, no spill to slow system RAM.
24 GB usable. Holds a 32B model at Q4 quality — dedicated vram memory, no spill to slow system RAM.
24 GB usable. Holds a 32B model at Q4 quality — dedicated vram memory, no spill to slow system RAM.
24 GB usable. Holds a 32B model at Q4 quality — dedicated vram memory, no spill to slow system RAM.
16 GB usable. Holds a 14B model at Q4 quality — dedicated vram memory, no spill to slow system RAM.
16 GB usable. Holds a 14B model at Q4 quality — dedicated vram memory, no spill to slow system RAM.
16 GB usable. Holds a 14B model at Q4 quality — dedicated vram memory, no spill to slow system RAM.
12 GB usable. Holds a 14B model at Q4 quality — dedicated vram memory, no spill to slow system RAM.
Pick up to three. The winner of each row is marked; the answer underneath says why, starting with the memory feature and then the rest of the device.
Sorted by usable model memory — the feature that decides which models a device can run at all.
| # | Device | Usable | Largest at Q4 | Price |
|---|---|---|---|---|
| 1 | GeForce RTX 5090 (desktop card) | 32 GB | 32B | $1,999 |
| 2 | GeForce RTX 4090 (desktop card) | 24 GB | 32B | — |
| 3 | Radeon RX 7900 XTX (desktop card) | 24 GB | 32B | — |
| 4 | GeForce RTX 3090 (desktop card, used market) | 24 GB | 32B | — |
| 5 | GeForce RTX 5070 Ti (desktop card) | 16 GB | 14B | $749 |
| 6 | GeForce RTX 5080 (desktop card) | 16 GB | 14B | $999 |
| 7 | GeForce RTX 5060 Ti 16GB (desktop card) | 16 GB | 14B | — |
| 8 | GeForce RTX 5070 (desktop card) | 12 GB | 14B | $549 |
Usable memory is what is left for the model itself. A graphics card uses all of its VRAM; a unified-memory device shares memory with the system, so about 75 percent of it can hold a model. The bar shows the selected model against that usable memory.
96 GB usable. Holds a 70B model at Q4 quality — unified memory, no spill to slow system RAM.
96 GB usable. Holds a 70B model at Q4 quality — unified memory, no spill to slow system RAM.
48 GB usable. Holds a 70B model at Q4 quality — unified memory, no spill to slow system RAM.
36 GB usable. Holds a 32B model at Q4 quality — unified memory, no spill to slow system RAM.
18 GB usable. Holds a 14B model at Q4 quality — unified memory, no spill to slow system RAM.
Pick up to three. The winner of each row is marked; the answer underneath says why, starting with the memory feature and then the rest of the device.
Sorted by usable model memory — the feature that decides which models a device can run at all.
| # | Device | Usable | Largest at Q4 | Price |
|---|---|---|---|---|
| 1 | Framework Desktop — Ryzen AI Max+ 395, 128GB | 96 GB | 70B | $3,449 |
| 2 | NVIDIA DGX Spark — GB10, 128GB | 96 GB | 70B | $4,699 |
| 3 | Framework Desktop — Ryzen AI Max+ 395, 64GB | 48 GB | 70B | $1,959 |
| 4 | Mac mini — M4 Pro, 48GB | 36 GB | 32B | — |
| 5 | Mac mini — M4, 24GB | 18 GB | 14B | — |
ModelFit. Laptops, desktop GPUs and mini PCs compared by one feature first: the memory your model can actually use — then everything else.
Specifications from manufacturer listings. Prices from retailer listings; check the retailer for the current price.