Gemma 4 12B at 4-bit (Q4_K_M) with 8k of context needs 8.2 GB on any machine.
Smallest that runs it well
16 GB
RTX 5070 Ti PC, RTX 5060 Ti 16 GB PC. It needs 8.2 GB.
RTX 5090 PC
At most 180 tokens/s, estimated from memory bandwidth for one user.
$0.17 an hour
One RTX 3090 on SaladCloud, home PCs, on demand before tax, or $0.44 for one L4 on Jarvislabs, which runs its own machines.
The answers above are for every machine; the list shows the kind you picked.
Runs it well 16 GB for the GPU · Windows or Linux
Yes
7.0 GB left
The same 16 GB as the RTX 5060 Ti 16 GB, read about twice as fast.
Where to buy: NVIDIA Marketplace
16 GB for the GPU · Windows or Linux
Yes
7.0 GB left
Buy the 16 GB version: the 8 GB card shares the name and runs far less.
Where to buy: NVIDIA Marketplace
24 GB for the GPU · Windows or Linux
≤ 101 tokens/s
Needs 8.2 GB
Yes
14.6 GB left
Sold mostly used now. The same 24 GB as a used RTX 3090, read a little faster.
24 GB for the GPU · Windows or Linux
Yes
14.6 GB left
Sold used only. Check the seller, the fans and the warranty before paying.
24 GB for the GPU · Olares OS
Yes
14.6 GB left
A 3.5-litre box with the laptop version of the RTX 5090 and 96 GB of system memory. Models must fit in the GPU's 24 GB; the system memory does not add to it.
Where to buy: Olares
24 GB for the GPU · Windows or Linux
Yes
14.6 GB left
Fewer local AI tools are tested on Intel cards than on NVIDIA or AMD ones. Check yours before buying.
Where to buy: Intel's list of shops
24 GB for the GPU · Ghost's own software
Yes
14.6 GB left
A small always-on box with a 24 GB RTX PRO 4000 Blackwell SFF Edition, sold with its own assistant software.
Where to buy: Ghost
32 GB for the GPU · Windows or Linux
≤ 180 tokens/s
Needs 8.2 GB
Yes
22.2 GB left
The fastest card you can put in a normal PC. It needs a large power supply and case.
Where to buy: NVIDIA Marketplace
32 GB for the GPU · Windows or Linux
Yes
22.2 GB left
AMD runs llama.cpp and Ollama well; image and video tools mostly expect NVIDIA.
Where to buy: AMD's list of shops
48 GB for the GPU · macOS
Yes
37.4 GB left
Choose 64 GB when you order: Mac memory cannot be added later.
Where to buy: Apple
96 GB for the GPU · Windows or Linux
≤ 180 tokens/s
Needs 8.2 GB
Yes
83.0 GB left
The most memory on one card, read as fast as an RTX 5090. Sold mostly through workstation makers and resellers.
Where to buy: NVIDIA Marketplace
96 GB for the GPU · macOS
Yes
83.0 GB left
128 GB needs the 40-core GPU version, chosen when you order.
Where to buy: Apple
96 GB for the GPU · Windows or Linux
Yes
83.0 GB left
102 GB for the GPU · Windows 11 (Arm)
Yes
88.7 GB left
Only the 128 GB versions hold the big models. Windows on Arm: check your apps run on it.
Where to buy: NVIDIA Marketplace
126 GB for the GPU · DGX OS (Linux)
Yes
111.5 GB left
NVIDIA's and its partners' GB10 boxes share the chip and the 128 GB. Big models load; dense ones answer slowly.
Where to buy: NVIDIA · Dell · Lenovo · Acer
384 GB for the GPU · macOS
≤ 121 tokens/s
Needs 8.2 GB
Yes
356.6 GB left
Shown at 512 GB, which needs the 80-core GPU chip; 96 and 256 GB versions run less. Long prompts are slow to read in on any Mac.
Where to buy: Apple
No machine of this kind is in the list.
Qwen3.8 27B at 4-bit (Q4_K_M) with 8k of context needs 18.0 GB on any machine.
Smallest that runs it well
24 GB
RTX 4090 PC, RTX 3090 PC (used), Olares One. It needs 18.0 GB.
RTX 5090 PC
About 78 tokens/s, estimated from memory bandwidth for one user.
$0.17 an hour
One RTX 3090 on SaladCloud, home PCs, on demand before tax, or $0.50 on RunPod, which runs its own machines.
The answers above are for every machine; the list shows the kind you picked.
Runs it well 24 GB for the GPU · Windows or Linux
44 tokens/s
Needs 18.0 GB
Yes
4.8 GB left
Sold mostly used now. The same 24 GB as a used RTX 3090, read a little faster.
24 GB for the GPU · Windows or Linux
41 tokens/s
Needs 18.0 GB
Yes
4.8 GB left
Sold used only. Check the seller, the fans and the warranty before paying.
24 GB for the GPU · Olares OS
39 tokens/s
Needs 18.0 GB
Yes
4.8 GB left
A 3.5-litre box with the laptop version of the RTX 5090 and 96 GB of system memory. Models must fit in the GPU's 24 GB; the system memory does not add to it.
Where to buy: Olares
32 GB for the GPU · Windows or Linux
78 tokens/s
Needs 18.0 GB
Yes
12.4 GB left
The fastest card you can put in a normal PC. It needs a large power supply and case.
Where to buy: NVIDIA Marketplace
32 GB for the GPU · Windows or Linux
28 tokens/s
Needs 18.0 GB
Yes
12.4 GB left
AMD runs llama.cpp and Ollama well; image and video tools mostly expect NVIDIA.
Where to buy: AMD's list of shops
96 GB for the GPU · Windows or Linux
78 tokens/s
Needs 18.0 GB
Yes
73.2 GB left
The most memory on one card, read as fast as an RTX 5090. Sold mostly through workstation makers and resellers.
Where to buy: NVIDIA Marketplace
96 GB for the GPU · macOS
27 tokens/s
Needs 18.0 GB
Yes
73.2 GB left
128 GB needs the 40-core GPU version, chosen when you order.
Where to buy: Apple
384 GB for the GPU · macOS
52 tokens/s
Needs 18.0 GB
Yes
346.8 GB left
Shown at 512 GB, which needs the 80-core GPU chip; 96 and 256 GB versions run less. Long prompts are slow to read in on any Mac.
Where to buy: Apple
Runs it, with little room or slowly 24 GB for the GPU · Windows or Linux
20 tokens/s
Needs 18.0 GB
Yes
4.8 GB left
Fewer local AI tools are tested on Intel cards than on NVIDIA or AMD ones. Check yours before buying.
Where to buy: Intel's list of shops
24 GB for the GPU · Ghost's own software
19 tokens/s
Needs 18.0 GB
Yes
4.8 GB left
A small always-on box with a 24 GB RTX PRO 4000 Blackwell SFF Edition, sold with its own assistant software.
Where to buy: Ghost
48 GB for the GPU · macOS
13 tokens/s
Needs 18.0 GB
Yes
27.6 GB left
Choose 64 GB when you order: Mac memory cannot be added later.
Where to buy: Apple
96 GB for the GPU · Windows or Linux
11 tokens/s
Needs 18.0 GB
Yes
73.2 GB left
102 GB for the GPU · Windows 11 (Arm)
13 tokens/s
Needs 18.0 GB
Yes
78.9 GB left
Only the 128 GB versions hold the big models. Windows on Arm: check your apps run on it.
Where to buy: NVIDIA Marketplace
126 GB for the GPU · DGX OS (Linux)
12 tokens/s
Needs 18.0 GB
Yes
101.7 GB left
NVIDIA's and its partners' GB10 boxes share the chip and the 128 GB. Big models load; dense ones answer slowly.
Where to buy: NVIDIA · Dell · Lenovo · Acer
Does not hold it 16 GB for the GPU · Windows or Linux
No
Short by 2.8 GB
The same 16 GB as the RTX 5060 Ti 16 GB, read about twice as fast.
Where to buy: NVIDIA Marketplace
16 GB for the GPU · Windows or Linux
No
Short by 2.8 GB
Buy the 16 GB version: the 8 GB card shares the name and runs far less.
Where to buy: NVIDIA Marketplace
No machine of this kind is in the list.
gpt-oss 120B at 4-bit (Q4_K_M) with 8k of context needs 71.4 GB on any machine.
Smallest that runs it well
96 GB
RTX PRO 6000 workstation, MacBook Pro M5 Max, Ryzen AI Max+ 395 mini PC. It needs 71.4 GB.
RTX PRO 6000 workstation
At most 424 tokens/s, estimated from memory bandwidth for one user.
$1.69 an hour
One RTX PRO 6000 on RunPod Community, third-party host, on demand before tax, or $1.80 on Nebius, which runs its own machines.
The answers above are for every machine; the list shows the kind you picked.
Runs it well 96 GB for the GPU · Windows or Linux
≤ 424 tokens/s
Needs 71.4 GB
Yes
19.8 GB left
The most memory on one card, read as fast as an RTX 5090. Sold mostly through workstation makers and resellers.
Where to buy: NVIDIA Marketplace
96 GB for the GPU · macOS
≤ 145 tokens/s
Needs 71.4 GB
Yes
19.8 GB left
128 GB needs the 40-core GPU version, chosen when you order.
Where to buy: Apple
96 GB for the GPU · Windows or Linux
61 tokens/s
Needs 71.4 GB
Yes
19.8 GB left
102 GB for the GPU · Windows 11 (Arm)
71 tokens/s
Needs 71.4 GB
Yes
25.5 GB left
Only the 128 GB versions hold the big models. Windows on Arm: check your apps run on it.
Where to buy: NVIDIA Marketplace
126 GB for the GPU · DGX OS (Linux)
65 tokens/s
Needs 71.4 GB
Yes
48.3 GB left
NVIDIA's and its partners' GB10 boxes share the chip and the 128 GB. Big models load; dense ones answer slowly.
Where to buy: NVIDIA · Dell · Lenovo · Acer
384 GB for the GPU · macOS
≤ 284 tokens/s
Needs 71.4 GB
Yes
293.4 GB left
Shown at 512 GB, which needs the 80-core GPU chip; 96 and 256 GB versions run less. Long prompts are slow to read in on any Mac.
Where to buy: Apple
Does not hold it 16 GB for the GPU · Windows or Linux
No
Short by 56.2 GB
The same 16 GB as the RTX 5060 Ti 16 GB, read about twice as fast.
Where to buy: NVIDIA Marketplace
16 GB for the GPU · Windows or Linux
No
Short by 56.2 GB
Buy the 16 GB version: the 8 GB card shares the name and runs far less.
Where to buy: NVIDIA Marketplace
24 GB for the GPU · Windows or Linux
No
Short by 48.6 GB
Sold mostly used now. The same 24 GB as a used RTX 3090, read a little faster.
24 GB for the GPU · Windows or Linux
No
Short by 48.6 GB
Sold used only. Check the seller, the fans and the warranty before paying.
24 GB for the GPU · Olares OS
No
Short by 48.6 GB
A 3.5-litre box with the laptop version of the RTX 5090 and 96 GB of system memory. Models must fit in the GPU's 24 GB; the system memory does not add to it.
Where to buy: Olares
24 GB for the GPU · Windows or Linux
No
Short by 48.6 GB
Fewer local AI tools are tested on Intel cards than on NVIDIA or AMD ones. Check yours before buying.
Where to buy: Intel's list of shops
24 GB for the GPU · Ghost's own software
No
Short by 48.6 GB
A small always-on box with a 24 GB RTX PRO 4000 Blackwell SFF Edition, sold with its own assistant software.
Where to buy: Ghost
32 GB for the GPU · Windows or Linux
No
Short by 41.0 GB
The fastest card you can put in a normal PC. It needs a large power supply and case.
Where to buy: NVIDIA Marketplace
32 GB for the GPU · Windows or Linux
No
Short by 41.0 GB
AMD runs llama.cpp and Ollama well; image and video tools mostly expect NVIDIA.
Where to buy: AMD's list of shops
48 GB for the GPU · macOS
No
Short by 25.8 GB
Choose 64 GB when you order: Mac memory cannot be added later.
Where to buy: Apple
No machine of this kind is in the list.
Gemma 4 12B at 4-bit (Q4_K_M) with 64k of context, the context agents ask for, needs 9.1 GB on any machine.
Smallest that runs it well
16 GB
RTX 5070 Ti PC, RTX 5060 Ti 16 GB PC. It needs 9.1 GB.
RTX 5090 PC
At most 180 tokens/s, estimated from memory bandwidth for one user.
$0.17 an hour
One RTX 3090 on SaladCloud, home PCs, on demand before tax, or $0.44 for one L4 on Jarvislabs, which runs its own machines.
The answers above are for every machine; the list shows the kind you picked.
Runs it well 16 GB for the GPU · Windows or Linux
Yes
6.1 GB left
The same 16 GB as the RTX 5060 Ti 16 GB, read about twice as fast.
Where to buy: NVIDIA Marketplace
16 GB for the GPU · Windows or Linux
Yes
6.1 GB left
Buy the 16 GB version: the 8 GB card shares the name and runs far less.
Where to buy: NVIDIA Marketplace
24 GB for the GPU · Windows or Linux
≤ 101 tokens/s
Needs 9.1 GB
Yes
13.7 GB left
Sold mostly used now. The same 24 GB as a used RTX 3090, read a little faster.
24 GB for the GPU · Windows or Linux
Yes
13.7 GB left
Sold used only. Check the seller, the fans and the warranty before paying.
24 GB for the GPU · Olares OS
Yes
13.7 GB left
A 3.5-litre box with the laptop version of the RTX 5090 and 96 GB of system memory. Models must fit in the GPU's 24 GB; the system memory does not add to it.
Where to buy: Olares
24 GB for the GPU · Windows or Linux
Yes
13.7 GB left
Fewer local AI tools are tested on Intel cards than on NVIDIA or AMD ones. Check yours before buying.
Where to buy: Intel's list of shops
24 GB for the GPU · Ghost's own software
Yes
13.7 GB left
A small always-on box with a 24 GB RTX PRO 4000 Blackwell SFF Edition, sold with its own assistant software.
Where to buy: Ghost
32 GB for the GPU · Windows or Linux
≤ 180 tokens/s
Needs 9.1 GB
Yes
21.3 GB left
The fastest card you can put in a normal PC. It needs a large power supply and case.
Where to buy: NVIDIA Marketplace
32 GB for the GPU · Windows or Linux
Yes
21.3 GB left
AMD runs llama.cpp and Ollama well; image and video tools mostly expect NVIDIA.
Where to buy: AMD's list of shops
48 GB for the GPU · macOS
Yes
36.5 GB left
Choose 64 GB when you order: Mac memory cannot be added later.
Where to buy: Apple
96 GB for the GPU · Windows or Linux
≤ 180 tokens/s
Needs 9.1 GB
Yes
82.1 GB left
The most memory on one card, read as fast as an RTX 5090. Sold mostly through workstation makers and resellers.
Where to buy: NVIDIA Marketplace
96 GB for the GPU · macOS
Yes
82.1 GB left
128 GB needs the 40-core GPU version, chosen when you order.
Where to buy: Apple
96 GB for the GPU · Windows or Linux
Yes
82.1 GB left
102 GB for the GPU · Windows 11 (Arm)
Yes
87.8 GB left
Only the 128 GB versions hold the big models. Windows on Arm: check your apps run on it.
Where to buy: NVIDIA Marketplace
126 GB for the GPU · DGX OS (Linux)
Yes
110.6 GB left
NVIDIA's and its partners' GB10 boxes share the chip and the 128 GB. Big models load; dense ones answer slowly.
Where to buy: NVIDIA · Dell · Lenovo · Acer
384 GB for the GPU · macOS
≤ 121 tokens/s
Needs 9.1 GB
Yes
355.7 GB left
Shown at 512 GB, which needs the 80-core GPU chip; 96 and 256 GB versions run less. Long prompts are slow to read in on any Mac.
Where to buy: Apple
No machine of this kind is in the list.
Qwen3.8 27B at 4-bit (Q4_K_M) with 64k of context, the context agents ask for, needs 21.7 GB on any machine.
Smallest that runs it well
32 GB
RTX 5090 PC, Radeon AI PRO R9700 PC. It needs 21.7 GB.
RTX 5090 PC
About 78 tokens/s, estimated from memory bandwidth for one user.
$0.33 an hour
One RTX A6000 on RunPod Community, third-party host, on demand before tax, or $0.50 on Hyperstack, which runs its own machines.
The answers above are for every machine; the list shows the kind you picked.
Runs it well 32 GB for the GPU · Windows or Linux
78 tokens/s
Needs 21.7 GB
Yes
8.7 GB left
The fastest card you can put in a normal PC. It needs a large power supply and case.
Where to buy: NVIDIA Marketplace
32 GB for the GPU · Windows or Linux
28 tokens/s
Needs 21.7 GB
Yes
8.7 GB left
AMD runs llama.cpp and Ollama well; image and video tools mostly expect NVIDIA.
Where to buy: AMD's list of shops
96 GB for the GPU · Windows or Linux
78 tokens/s
Needs 21.7 GB
Yes
69.5 GB left
The most memory on one card, read as fast as an RTX 5090. Sold mostly through workstation makers and resellers.
Where to buy: NVIDIA Marketplace
96 GB for the GPU · macOS
27 tokens/s
Needs 21.7 GB
Yes
69.5 GB left
128 GB needs the 40-core GPU version, chosen when you order.
Where to buy: Apple
384 GB for the GPU · macOS
52 tokens/s
Needs 21.7 GB
Yes
343.1 GB left
Shown at 512 GB, which needs the 80-core GPU chip; 96 and 256 GB versions run less. Long prompts are slow to read in on any Mac.
Where to buy: Apple
Runs it, with little room or slowly 48 GB for the GPU · macOS
13 tokens/s
Needs 21.7 GB
Yes
23.9 GB left
Choose 64 GB when you order: Mac memory cannot be added later.
Where to buy: Apple
96 GB for the GPU · Windows or Linux
11 tokens/s
Needs 21.7 GB
Yes
69.5 GB left
102 GB for the GPU · Windows 11 (Arm)
13 tokens/s
Needs 21.7 GB
Yes
75.2 GB left
Only the 128 GB versions hold the big models. Windows on Arm: check your apps run on it.
Where to buy: NVIDIA Marketplace
126 GB for the GPU · DGX OS (Linux)
12 tokens/s
Needs 21.7 GB
Yes
98.0 GB left
NVIDIA's and its partners' GB10 boxes share the chip and the 128 GB. Big models load; dense ones answer slowly.
Where to buy: NVIDIA · Dell · Lenovo · Acer
24 GB for the GPU · Windows or Linux
44 tokens/s
Needs 21.7 GB
Tight
1.1 GB left
Sold mostly used now. The same 24 GB as a used RTX 3090, read a little faster.
24 GB for the GPU · Windows or Linux
41 tokens/s
Needs 21.7 GB
Tight
1.1 GB left
Sold used only. Check the seller, the fans and the warranty before paying.
24 GB for the GPU · Olares OS
39 tokens/s
Needs 21.7 GB
Tight
1.1 GB left
A 3.5-litre box with the laptop version of the RTX 5090 and 96 GB of system memory. Models must fit in the GPU's 24 GB; the system memory does not add to it.
Where to buy: Olares
24 GB for the GPU · Windows or Linux
20 tokens/s
Needs 21.7 GB
Tight
1.1 GB left
Fewer local AI tools are tested on Intel cards than on NVIDIA or AMD ones. Check yours before buying.
Where to buy: Intel's list of shops
24 GB for the GPU · Ghost's own software
19 tokens/s
Needs 21.7 GB
Tight
1.1 GB left
A small always-on box with a 24 GB RTX PRO 4000 Blackwell SFF Edition, sold with its own assistant software.
Where to buy: Ghost
Does not hold it 16 GB for the GPU · Windows or Linux
No
Short by 6.5 GB
The same 16 GB as the RTX 5060 Ti 16 GB, read about twice as fast.
Where to buy: NVIDIA Marketplace
16 GB for the GPU · Windows or Linux
No
Short by 6.5 GB
Buy the 16 GB version: the 8 GB card shares the name and runs far less.
Where to buy: NVIDIA Marketplace
No machine of this kind is in the list.
Qwen3-Coder-Next 80B-A3B (MoE) at 4-bit (Q4_K_M) with 64k of context, the context agents ask for, needs 50.3 GB on any machine.
Smallest that runs it well
96 GB
RTX PRO 6000 workstation, MacBook Pro M5 Max, Ryzen AI Max+ 395 mini PC. It needs 50.3 GB.
RTX PRO 6000 workstation
At most 721 tokens/s, estimated from memory bandwidth for one user.
$1.35 an hour
One A100 80 GB on Hyperstack, on demand before tax.
The answers above are for every machine; the list shows the kind you picked.
Runs it well 96 GB for the GPU · Windows or Linux
≤ 721 tokens/s
Needs 50.3 GB
Yes
40.9 GB left
The most memory on one card, read as fast as an RTX 5090. Sold mostly through workstation makers and resellers.
Where to buy: NVIDIA Marketplace
96 GB for the GPU · macOS
≤ 247 tokens/s
Needs 50.3 GB
Yes
40.9 GB left
128 GB needs the 40-core GPU version, chosen when you order.
Where to buy: Apple
96 GB for the GPU · Windows or Linux
≤ 103 tokens/s
Needs 50.3 GB
Yes
40.9 GB left
102 GB for the GPU · Windows 11 (Arm)
≤ 121 tokens/s
Needs 50.3 GB
Yes
46.6 GB left
Only the 128 GB versions hold the big models. Windows on Arm: check your apps run on it.
Where to buy: NVIDIA Marketplace
126 GB for the GPU · DGX OS (Linux)
≤ 110 tokens/s
Needs 50.3 GB
Yes
69.4 GB left
NVIDIA's and its partners' GB10 boxes share the chip and the 128 GB. Big models load; dense ones answer slowly.
Where to buy: NVIDIA · Dell · Lenovo · Acer
384 GB for the GPU · macOS
≤ 483 tokens/s
Needs 50.3 GB
Yes
314.5 GB left
Shown at 512 GB, which needs the 80-core GPU chip; 96 and 256 GB versions run less. Long prompts are slow to read in on any Mac.
Where to buy: Apple
Does not hold it 16 GB for the GPU · Windows or Linux
No
Short by 35.1 GB
The same 16 GB as the RTX 5060 Ti 16 GB, read about twice as fast.
Where to buy: NVIDIA Marketplace
16 GB for the GPU · Windows or Linux
No
Short by 35.1 GB
Buy the 16 GB version: the 8 GB card shares the name and runs far less.
Where to buy: NVIDIA Marketplace
24 GB for the GPU · Windows or Linux
No
Short by 27.5 GB
Sold mostly used now. The same 24 GB as a used RTX 3090, read a little faster.
24 GB for the GPU · Windows or Linux
No
Short by 27.5 GB
Sold used only. Check the seller, the fans and the warranty before paying.
24 GB for the GPU · Olares OS
No
Short by 27.5 GB
A 3.5-litre box with the laptop version of the RTX 5090 and 96 GB of system memory. Models must fit in the GPU's 24 GB; the system memory does not add to it.
Where to buy: Olares
24 GB for the GPU · Windows or Linux
No
Short by 27.5 GB
Fewer local AI tools are tested on Intel cards than on NVIDIA or AMD ones. Check yours before buying.
Where to buy: Intel's list of shops
24 GB for the GPU · Ghost's own software
No
Short by 27.5 GB
A small always-on box with a 24 GB RTX PRO 4000 Blackwell SFF Edition, sold with its own assistant software.
Where to buy: Ghost
32 GB for the GPU · Windows or Linux
No
Short by 19.9 GB
The fastest card you can put in a normal PC. It needs a large power supply and case.
Where to buy: NVIDIA Marketplace
32 GB for the GPU · Windows or Linux
No
Short by 19.9 GB
AMD runs llama.cpp and Ollama well; image and video tools mostly expect NVIDIA.
Where to buy: AMD's list of shops
48 GB for the GPU · macOS
No
Short by 4.7 GB
Choose 64 GB when you order: Mac memory cannot be added later.
Where to buy: Apple
No machine of this kind is in the list.
Z-Image (Turbo), ComfyUI's own template, with the best file each card holds.
Smallest that runs it well
16 GB
RTX 5070 Ti PC, RTX 5060 Ti 16 GB PC.
10 of 10
Checked on the NVIDIA machines below, by memory. Image and video speed is not computed.
$0.17 an hour
One RTX 3090 on SaladCloud, home PCs, on demand before tax, or $0.44 for one L4 on Jarvislabs, which runs its own machines.
The answers above are for every machine; the list shows the kind you picked.
Holds it with room to work 16 GB for the GPU · Windows or Linux
Yes
2.5 GB left
The same 16 GB as the RTX 5060 Ti 16 GB, read about twice as fast.
Where to buy: NVIDIA Marketplace
16 GB for the GPU · Windows or Linux
Yes
2.5 GB left
Buy the 16 GB version: the 8 GB card shares the name and runs far less.
Where to buy: NVIDIA Marketplace
24 GB for the GPU · Windows or Linux
Yes
10.1 GB left
Sold mostly used now. The same 24 GB as a used RTX 3090, read a little faster.
24 GB for the GPU · Windows or Linux
Yes
10.1 GB left
Sold used only. Check the seller, the fans and the warranty before paying.
24 GB for the GPU · Olares OS
Yes
10.1 GB left
A 3.5-litre box with the laptop version of the RTX 5090 and 96 GB of system memory. Models must fit in the GPU's 24 GB; the system memory does not add to it.
Where to buy: Olares
24 GB for the GPU · Ghost's own software
Yes
10.1 GB left
A small always-on box with a 24 GB RTX PRO 4000 Blackwell SFF Edition, sold with its own assistant software.
Where to buy: Ghost
32 GB for the GPU · Windows or Linux
Yes
17.7 GB left
The fastest card you can put in a normal PC. It needs a large power supply and case.
Where to buy: NVIDIA Marketplace
96 GB for the GPU · Windows or Linux
Yes
78.5 GB left
The most memory on one card, read as fast as an RTX 5090. Sold mostly through workstation makers and resellers.
Where to buy: NVIDIA Marketplace
102 GB for the GPU · Windows 11 (Arm)
Yes
84.2 GB left
Only the 128 GB versions hold the big models. Windows on Arm: check your apps run on it.
Where to buy: NVIDIA Marketplace
126 GB for the GPU · DGX OS (Linux)
Yes
107.0 GB left
NVIDIA's and its partners' GB10 boxes share the chip and the 128 GB. Big models load; dense ones answer slowly.
Where to buy: NVIDIA · Dell · Lenovo · Acer
No machine of this kind is checked for images and video.
Not checked for images and video: Radeon AI PRO R9700 PC, Arc Pro B60 PC, Ryzen AI Max+ 395 mini PC, Mac mini M5 Pro, MacBook Pro M5 Max, Mac Studio M5 Ultra. They run ComfyUI too, but the memory check here covers NVIDIA cards only.
Krea 2 (Raw), ComfyUI's own template, with the best file each card holds.
Smallest that runs it well
16 GB
RTX 5070 Ti PC, RTX 5060 Ti 16 GB PC.
10 of 10
Checked on the NVIDIA machines below, by memory. Image and video speed is not computed.
$0.17 an hour
One RTX 3090 on SaladCloud, home PCs, on demand before tax, or $0.44 for one L4 on Jarvislabs, which runs its own machines.
The answers above are for every machine; the list shows the kind you picked.
Holds it with room to work 16 GB for the GPU · Windows or Linux
FP8 scaled file
Needs 13.4 GB
Yes
1.8 GB left
The same 16 GB as the RTX 5060 Ti 16 GB, read about twice as fast.
Where to buy: NVIDIA Marketplace
16 GB for the GPU · Windows or Linux
FP8 scaled file
Needs 13.4 GB
Yes
1.8 GB left
Buy the 16 GB version: the 8 GB card shares the name and runs far less.
Where to buy: NVIDIA Marketplace
24 GB for the GPU · Windows or Linux
FP8 scaled file
Needs 13.4 GB
Yes
9.4 GB left
Sold mostly used now. The same 24 GB as a used RTX 3090, read a little faster.
24 GB for the GPU · Windows or Linux
FP8 scaled file
Needs 13.4 GB
Yes
9.4 GB left
Sold used only. Check the seller, the fans and the warranty before paying.
24 GB for the GPU · Olares OS
FP8 scaled file
Needs 13.4 GB
Yes
9.4 GB left
A 3.5-litre box with the laptop version of the RTX 5090 and 96 GB of system memory. Models must fit in the GPU's 24 GB; the system memory does not add to it.
Where to buy: Olares
24 GB for the GPU · Ghost's own software
FP8 scaled file
Needs 13.4 GB
Yes
9.4 GB left
A small always-on box with a 24 GB RTX PRO 4000 Blackwell SFF Edition, sold with its own assistant software.
Where to buy: Ghost
32 GB for the GPU · Windows or Linux
Yes
3.9 GB left
The fastest card you can put in a normal PC. It needs a large power supply and case.
Where to buy: NVIDIA Marketplace
96 GB for the GPU · Windows or Linux
Yes
64.7 GB left
The most memory on one card, read as fast as an RTX 5090. Sold mostly through workstation makers and resellers.
Where to buy: NVIDIA Marketplace
102 GB for the GPU · Windows 11 (Arm)
Yes
70.4 GB left
Only the 128 GB versions hold the big models. Windows on Arm: check your apps run on it.
Where to buy: NVIDIA Marketplace
126 GB for the GPU · DGX OS (Linux)
Yes
93.2 GB left
NVIDIA's and its partners' GB10 boxes share the chip and the 128 GB. Big models load; dense ones answer slowly.
Where to buy: NVIDIA · Dell · Lenovo · Acer
No machine of this kind is checked for images and video.
Not checked for images and video: Radeon AI PRO R9700 PC, Arc Pro B60 PC, Ryzen AI Max+ 395 mini PC, Mac mini M5 Pro, MacBook Pro M5 Max, Mac Studio M5 Ultra. They run ComfyUI too, but the memory check here covers NVIDIA cards only.
FLUX.2 (dev), ComfyUI's own template, with the best file each card holds.
Smallest that runs it well
32 GB
RTX 5090 PC.
6 of 10
Checked on the NVIDIA machines below, by memory. Image and video speed is not computed.
$0.33 an hour
One RTX A6000 on RunPod Community, third-party host, on demand before tax, or $0.50 on Hyperstack, which runs its own machines.
The answers above are for every machine; the list shows the kind you picked.
Holds it with room to work 32 GB for the GPU · Windows or Linux
NVFP4 mixed file
Needs 23.1 GB
Yes
7.3 GB left
The fastest card you can put in a normal PC. It needs a large power supply and case.
Where to buy: NVIDIA Marketplace
96 GB for the GPU · Windows or Linux
Yes
26.4 GB left
The most memory on one card, read as fast as an RTX 5090. Sold mostly through workstation makers and resellers.
Where to buy: NVIDIA Marketplace
102 GB for the GPU · Windows 11 (Arm)
Yes
32.1 GB left
Only the 128 GB versions hold the big models. Windows on Arm: check your apps run on it.
Where to buy: NVIDIA Marketplace
126 GB for the GPU · DGX OS (Linux)
Yes
54.9 GB left
NVIDIA's and its partners' GB10 boxes share the chip and the 128 GB. Big models load; dense ones answer slowly.
Where to buy: NVIDIA · Dell · Lenovo · Acer
Holds it, only just 24 GB for the GPU · Olares OS
Tight
1.4 GB left
A 3.5-litre box with the laptop version of the RTX 5090 and 96 GB of system memory. Models must fit in the GPU's 24 GB; the system memory does not add to it.
Where to buy: Olares
24 GB for the GPU · Ghost's own software
Tight
1.4 GB left
A small always-on box with a 24 GB RTX PRO 4000 Blackwell SFF Edition, sold with its own assistant software.
Where to buy: Ghost
Streams it from system memory, slowly 16 GB for the GPU · Windows or Linux
Smallest file, NVFP4
Needs 21.4 GB
Streams
Short by 6.2 GB
The same 16 GB as the RTX 5060 Ti 16 GB, read about twice as fast.
Where to buy: NVIDIA Marketplace
16 GB for the GPU · Windows or Linux
Smallest file, NVFP4
Needs 21.4 GB
Streams
Short by 6.2 GB
Buy the 16 GB version: the 8 GB card shares the name and runs far less.
Where to buy: NVIDIA Marketplace
24 GB for the GPU · Windows or Linux
Smallest file, FP8 mixed
Needs 35.8 GB
Streams
Short by 13.0 GB
Sold mostly used now. The same 24 GB as a used RTX 3090, read a little faster.
24 GB for the GPU · Windows or Linux
Smallest file, FP8 mixed
Needs 35.8 GB
Streams
Short by 13.0 GB
Sold used only. Check the seller, the fans and the warranty before paying.
No machine of this kind is checked for images and video.
Not checked for images and video: Radeon AI PRO R9700 PC, Arc Pro B60 PC, Ryzen AI Max+ 395 mini PC, Mac mini M5 Pro, MacBook Pro M5 Max, Mac Studio M5 Ultra. They run ComfyUI too, but the memory check here covers NVIDIA cards only.
Wan 2.2 (Text and image to video, 5B), ComfyUI's own template, with the best file each card holds.
Smallest that runs it well
16 GB
RTX 5070 Ti PC, RTX 5060 Ti 16 GB PC.
10 of 10
Checked on the NVIDIA machines below, by memory. Image and video speed is not computed.
$0.17 an hour
One RTX 3090 on SaladCloud, home PCs, on demand before tax, or $0.44 for one L4 on Jarvislabs, which runs its own machines.
The answers above are for every machine; the list shows the kind you picked.
Holds it with room to work 16 GB for the GPU · Windows or Linux
Yes
3.8 GB left
The same 16 GB as the RTX 5060 Ti 16 GB, read about twice as fast.
Where to buy: NVIDIA Marketplace
16 GB for the GPU · Windows or Linux
Yes
3.8 GB left
Buy the 16 GB version: the 8 GB card shares the name and runs far less.
Where to buy: NVIDIA Marketplace
24 GB for the GPU · Windows or Linux
Yes
11.4 GB left
Sold mostly used now. The same 24 GB as a used RTX 3090, read a little faster.
24 GB for the GPU · Windows or Linux
Yes
11.4 GB left
Sold used only. Check the seller, the fans and the warranty before paying.
24 GB for the GPU · Olares OS
Yes
11.4 GB left
A 3.5-litre box with the laptop version of the RTX 5090 and 96 GB of system memory. Models must fit in the GPU's 24 GB; the system memory does not add to it.
Where to buy: Olares
24 GB for the GPU · Ghost's own software
Yes
11.4 GB left
A small always-on box with a 24 GB RTX PRO 4000 Blackwell SFF Edition, sold with its own assistant software.
Where to buy: Ghost
32 GB for the GPU · Windows or Linux
Yes
19.0 GB left
The fastest card you can put in a normal PC. It needs a large power supply and case.
Where to buy: NVIDIA Marketplace
96 GB for the GPU · Windows or Linux
Yes
79.8 GB left
The most memory on one card, read as fast as an RTX 5090. Sold mostly through workstation makers and resellers.
Where to buy: NVIDIA Marketplace
102 GB for the GPU · Windows 11 (Arm)
Yes
85.5 GB left
Only the 128 GB versions hold the big models. Windows on Arm: check your apps run on it.
Where to buy: NVIDIA Marketplace
126 GB for the GPU · DGX OS (Linux)
Yes
108.3 GB left
NVIDIA's and its partners' GB10 boxes share the chip and the 128 GB. Big models load; dense ones answer slowly.
Where to buy: NVIDIA · Dell · Lenovo · Acer
No machine of this kind is checked for images and video.
Not checked for images and video: Radeon AI PRO R9700 PC, Arc Pro B60 PC, Ryzen AI Max+ 395 mini PC, Mac mini M5 Pro, MacBook Pro M5 Max, Mac Studio M5 Ultra. They run ComfyUI too, but the memory check here covers NVIDIA cards only.
Wan 2.2 (Image to video, A14B), ComfyUI's own template, with the best file each card holds.
Smallest that runs it well
24 GB
RTX 4090 PC, RTX 3090 PC (used), Olares One, Ghost Core.
10 of 10
Checked on the NVIDIA machines below, by memory. Image and video speed is not computed.
$0.17 an hour
One RTX 3090 on SaladCloud, home PCs, on demand before tax, or $0.44 for one L4 on Jarvislabs, which runs its own machines.
The answers above are for every machine; the list shows the kind you picked.
Holds it with room to work 24 GB for the GPU · Windows or Linux
FP8 scaled file
Needs 14.5 GB
Yes
8.3 GB left
Sold mostly used now. The same 24 GB as a used RTX 3090, read a little faster.
24 GB for the GPU · Windows or Linux
FP8 scaled file
Needs 14.5 GB
Yes
8.3 GB left
Sold used only. Check the seller, the fans and the warranty before paying.
24 GB for the GPU · Olares OS
FP8 scaled file
Needs 14.5 GB
Yes
8.3 GB left
A 3.5-litre box with the laptop version of the RTX 5090 and 96 GB of system memory. Models must fit in the GPU's 24 GB; the system memory does not add to it.
Where to buy: Olares
24 GB for the GPU · Ghost's own software
FP8 scaled file
Needs 14.5 GB
Yes
8.3 GB left
A small always-on box with a 24 GB RTX PRO 4000 Blackwell SFF Edition, sold with its own assistant software.
Where to buy: Ghost
32 GB for the GPU · Windows or Linux
FP8 scaled file
Needs 14.5 GB
Yes
15.9 GB left
The fastest card you can put in a normal PC. It needs a large power supply and case.
Where to buy: NVIDIA Marketplace
96 GB for the GPU · Windows or Linux
Yes
62.4 GB left
The most memory on one card, read as fast as an RTX 5090. Sold mostly through workstation makers and resellers.
Where to buy: NVIDIA Marketplace
102 GB for the GPU · Windows 11 (Arm)
Yes
68.1 GB left
Only the 128 GB versions hold the big models. Windows on Arm: check your apps run on it.
Where to buy: NVIDIA Marketplace
126 GB for the GPU · DGX OS (Linux)
Yes
90.9 GB left
NVIDIA's and its partners' GB10 boxes share the chip and the 128 GB. Big models load; dense ones answer slowly.
Where to buy: NVIDIA · Dell · Lenovo · Acer
Holds it, only just 16 GB for the GPU · Windows or Linux
FP8 scaled file
Needs 14.5 GB
Tight
0.7 GB left
The same 16 GB as the RTX 5060 Ti 16 GB, read about twice as fast.
Where to buy: NVIDIA Marketplace
16 GB for the GPU · Windows or Linux
FP8 scaled file
Needs 14.5 GB
Tight
0.7 GB left
Buy the 16 GB version: the 8 GB card shares the name and runs far less.
Where to buy: NVIDIA Marketplace
No machine of this kind is checked for images and video.
Not checked for images and video: Radeon AI PRO R9700 PC, Arc Pro B60 PC, Ryzen AI Max+ 395 mini PC, Mac mini M5 Pro, MacBook Pro M5 Max, Mac Studio M5 Ultra. They run ComfyUI too, but the memory check here covers NVIDIA cards only.
LTX-2.5 (Distilled), ComfyUI's own template, with the best file each card holds.
Smallest that runs it well
32 GB
RTX 5090 PC.
6 of 10
Checked on the NVIDIA machines below, by memory. Image and video speed is not computed.
$0.33 an hour
One RTX A6000 on RunPod Community, third-party host, on demand before tax, or $0.50 on Hyperstack, which runs its own machines.
The answers above are for every machine; the list shows the kind you picked.
Holds it with room to work 32 GB for the GPU · Windows or Linux
INT8 ConvRot file
Needs 26.3 GB
Yes
4.1 GB left
The fastest card you can put in a normal PC. It needs a large power supply and case.
Where to buy: NVIDIA Marketplace
96 GB for the GPU · Windows or Linux
Yes
46.3 GB left
The most memory on one card, read as fast as an RTX 5090. Sold mostly through workstation makers and resellers.
Where to buy: NVIDIA Marketplace
102 GB for the GPU · Windows 11 (Arm)
Yes
52.0 GB left
Only the 128 GB versions hold the big models. Windows on Arm: check your apps run on it.
Where to buy: NVIDIA Marketplace
126 GB for the GPU · DGX OS (Linux)
Yes
74.8 GB left
NVIDIA's and its partners' GB10 boxes share the chip and the 128 GB. Big models load; dense ones answer slowly.
Where to buy: NVIDIA · Dell · Lenovo · Acer
Holds it, only just 24 GB for the GPU · Olares OS
Tight
1.3 GB left
A 3.5-litre box with the laptop version of the RTX 5090 and 96 GB of system memory. Models must fit in the GPU's 24 GB; the system memory does not add to it.
Where to buy: Olares
24 GB for the GPU · Ghost's own software
Tight
1.3 GB left
A small always-on box with a 24 GB RTX PRO 4000 Blackwell SFF Edition, sold with its own assistant software.
Where to buy: Ghost
Streams it from system memory, slowly 16 GB for the GPU · Windows or Linux
Smallest file, NVFP4
Needs 21.5 GB
Streams
Short by 6.3 GB
The same 16 GB as the RTX 5060 Ti 16 GB, read about twice as fast.
Where to buy: NVIDIA Marketplace
16 GB for the GPU · Windows or Linux
Smallest file, NVFP4
Needs 21.5 GB
Streams
Short by 6.3 GB
Buy the 16 GB version: the 8 GB card shares the name and runs far less.
Where to buy: NVIDIA Marketplace
24 GB for the GPU · Windows or Linux
Smallest file, INT8 ConvRot
Needs 24.3 GB
Streams
Short by 1.5 GB
Sold mostly used now. The same 24 GB as a used RTX 3090, read a little faster.
24 GB for the GPU · Windows or Linux
Smallest file, INT8 ConvRot
Needs 24.3 GB
Streams
Short by 1.5 GB
Sold used only. Check the seller, the fans and the warranty before paying.
No machine of this kind is checked for images and video.
Not checked for images and video: Radeon AI PRO R9700 PC, Arc Pro B60 PC, Ryzen AI Max+ 395 mini PC, Mac mini M5 Pro, MacBook Pro M5 Max, Mac Studio M5 Ultra. They run ComfyUI too, but the memory check here covers NVIDIA cards only.
Choosing for a team? Private AI for your team sizes one machine by how many people ask at once.