readygpu

Mac Studio M3 Ultra 256GB vs NVIDIA GeForce RTX 5090

For running models locally, priced in the US. At 32k context the first runs 75 models fully in memory, the second 52.

In short

The Mac Studio M3 Ultra 256GB has more memory for models: 256 GB against 32 GB. At 32k context the Mac Studio M3 Ultra 256GB runs 75 of the models we track fully in memory and the NVIDIA GeForce RTX 5090 52; only the Mac Studio M3 Ultra 256GB holds DeepSeek V4 Flash. On the models both run fully in memory, the NVIDIA GeForce RTX 5090 is about 2.1× as fast as the Mac Studio M3 Ultra 256GB (estimated, median across models). In the US the Mac Studio M3 Ultra 256GB costs $7,999 at Amazon and the NVIDIA GeForce RTX 5090 $7,489 at Amazon: the NVIDIA GeForce RTX 5090 is $510 cheaper.

Updated . Speeds are estimates; how they are worked out.

Mac Studio M3 Ultra 256GB

$7,999

Amazon

Add to build
NVIDIA GeForce RTX 5090

$7,489

Amazon

Add to build

Specifications

SpecMac Studio M3 Ultra 256GBNVIDIA GeForce RTX 5090
ChipApple M3 Ultra (80-core GPU)–
Memory bus1024-bit512-bit
Memory bandwidth819 GB/s1,792 GB/s
FP16 compute28 TFLOPS209 TFLOPS
SoftwareMetalCUDA
Unified memory256 GB–
GPU can use by default192 GB–
GPU can use, tuned248 GB–
Operating systemmacos–
ProcessorM3 Ultra 32-core–
Storage2000 GB–
GPU–GeForce RTX 5090
Memory–32 GB GDDR7
Board power–575 W
PCIe generation–Gen 5
PCIe lanes–x16

What each runs

Estimated decode speed with the most precise quant that fits each one est.

ModelMac Studio M3 Ultra 256GBNVIDIA GeForce RTX 5090
DeepSeek V4 Flash23tok/sMXFP4won't run
GLM 5.3 Flash20tok/sUD-Q4_K_XLwon't run
Qwen 3.8 Flash Next5tok/sQ8_0won't run
MiniMax M2.718tok/sUD-Q6_Kwon't run
Qwen 3.8 27B17tok/sQ8_035tok/sQ8_0
Qwen 3.5 122B-A10B28tok/sQ8_0offload
Qwen 3.6 27B17tok/sQ8_036tok/sQ8_0
gpt-oss 120B74tok/sMXFP4won't run
Qwen3 Coder Next64tok/sQ8_0offload
Nemotron 3 Super 120B-A12B25tok/sQ8_0offload
Gemma 4 31B15tok/sQ8_039tok/sQ6_K
Mistral Small 4 119B39tok/sQ8_0offload
Qwen 3.6 35B-A3B84tok/sQ8_0201tok/sQ6_K
Qwen 3.5 27B17tok/sQ8_036tok/sQ8_0
Gemma 4 26B-A4B63tok/sQ8_0132tok/sQ8_0
GLM 4.5 Air13tok/sQ8_0offload
Qwen 3.5 35B-A3B84tok/sQ8_0212tok/sQ6_K
Qwen3 Next 80B-A3B80tok/sQ8_0offload
GLM 4.7 Flash45tok/sQ8_0114tok/sQ6_K
Devstral Small 2 24B18tok/sQ8_037tok/sQ8_0
Nemotron 3.5 Lightning 30B-A3B89tok/sUD-Q6_K_XLoffload
gpt-oss 20B99tok/sMXFP4206tok/sMXFP4
Kimi Linear 48B-A3B91tok/sQ8_0311tok/sQ4_K_M
Qwen3 Coder 30B-A3B50tok/sQ8_0117tok/sQ6_K
Qwen3 32B12tok/sQ8_035tok/sQ5_K_M
Gemma 4 12B39tok/sQ8_082tok/sQ8_0
Qwen 3.5 9B50tok/sQ8_0104tok/sQ8_0
Mistral Small 3.2 24B18tok/sQ8_037tok/sQ8_0
Gemma 3 27B17tok/sQ8_035tok/sQ8_0
DeepSeek R1 Distill Qwen 32B12tok/sQ8_035tok/sQ5_K_M
Llama 3.3 70B6tok/sQ8_0offload
Ministral 3 14B27tok/sQ8_056tok/sQ8_0
Qwen3 14B25tok/sQ8_053tok/sQ8_0
DeepSeek R1 0528 Qwen3 8B39tok/sQ8_082tok/sQ8_0
Qwen3 8B39tok/sQ8_082tok/sQ8_0
Phi-4 14B24tok/sQ8_050tok/sQ8_0
Qwen 3.5 4B95tok/sQ8_0198tok/sQ8_0
Ministral 3 8B39tok/sQ8_082tok/sQ8_0
Granite 4.2 8B36tok/sQ8_076tok/sQ8_0
Gemma 4 E4B58tok/sQ8_0121tok/sQ8_0
LFM2.5 8B-A1B119tok/sQ8_0249tok/sQ8_0
Llama 3.1 8B49tok/sQ6_K102tok/sQ6_K
Qwen2.5 0.5B570tok/sQ8_01190tok/sQ8_0
Qwen3.5 0.8B432tok/sQ8_0901tok/sQ8_0
Qwen3.5 2B287tok/sQ5_K_M598tok/sQ5_K_M
DeepSeek R1 Distill Qwen 1.5B188tok/sQ8_0392tok/sQ8_0
MiniCPM5 2B130tok/sQ8_0272tok/sQ8_0
LFM2.5 2.6B109tok/sQ8_0227tok/sQ8_0
LFM2.5 VL 3B109tok/sQ8_0227tok/sQ8_0
Mellum2.1 12B A2.5B Thinking101tok/sQ8_0210tok/sQ8_0
granite vision 4.1 4b84tok/sQ8_0176tok/sQ8_0
granite 4.2 3b81tok/sQ8_0169tok/sQ8_0
Kimi VL A3B75tok/smlx-8bit156tok/smlx-8bit
LFM2.5 24B A2B66tok/sQ8_0139tok/sQ8_0
DeepSeek R1 Distill Qwen 7B53tok/sQ8_0111tok/sQ8_0
GLM 4.6V Flash47tok/sQ8_098tok/sQ8_0
DeepSeek R1 Distill Llama 8B42tok/sQ8_087tok/sQ8_0
GLM 4.1V 9B Thinking41tok/smlx-8bit85tok/smlx-8bit
Ministral 3 8B Reasoning 251239tok/sQ8_082tok/sQ8_0
DeepSeek R1 Distill Qwen 14B24tok/sQ8_050tok/sQ8_0
MiniMax M218tok/sQ6_Kwon't run
MiniMax M2.118tok/sQ6_Kwon't run
MiniMax M2.518tok/sQ6_Kwon't run
Devstral Small 250718tok/sQ8_037tok/sQ8_0
Magistral Small 250618tok/sQ8_037tok/sQ8_0
Magistral Small 250718tok/sQ8_037tok/sQ8_0
Magistral Small 250918tok/sQ8_037tok/sQ8_0
DeepSeek V4 Flash 073116tok/sMXFP4won't run
GLM 4.5V15tok/sQ8_0won't run
GLM 4.6V15tok/sQ8_0won't run
DeepSeek V4 Flash Vision Exp15tok/sMXFP4won't run
granite 4.2 30b13tok/sQ8_038tok/sQ5_K_M
DeepSeek R1 Distill Llama 70B6tok/sQ8_0offload
Devstral 2 123B Instruct 25124tok/sQ8_0won't run
Mistral Medium 3.5 128B4tok/sQ8_0won't run
DeepSeek R1won't runwon't run
DeepSeek R1 0528won't runwon't run
DeepSeek V3 0324won't runwon't run
DeepSeek V3.1won't runwon't run
DeepSeek V3.2won't runwon't run
GLM 4.7won't runwon't run
GLM 5.1won't runwon't run
GLM 5.2won't runwon't run
GLM 5.3won't runwon't run
Kimi K2won't runwon't run
Kimi K2 Instruct 0905won't runwon't run
Kimi K2 Thinkingwon't runwon't run
Kimi K2.5won't runwon't run
MiniMax M3won't runwon't run

Questions

Which is better for local AI, the Mac Studio M3 Ultra 256GB or the NVIDIA GeForce RTX 5090?
Choose the Mac Studio M3 Ultra 256GB to run bigger models. Choose the NVIDIA GeForce RTX 5090 for speed. Choose the NVIDIA GeForce RTX 5090 to spend less.
Which can run bigger models, the Mac Studio M3 Ultra 256GB or the NVIDIA GeForce RTX 5090?
At 32k context the Mac Studio M3 Ultra 256GB runs 75 of the models we track fully in memory and the NVIDIA GeForce RTX 5090 52; only the Mac Studio M3 Ultra 256GB holds DeepSeek V4 Flash.
Which is faster for local AI, the Mac Studio M3 Ultra 256GB or the NVIDIA GeForce RTX 5090?
On the models both run fully in memory, the NVIDIA GeForce RTX 5090 is about 2.1× as fast as the Mac Studio M3 Ultra 256GB (estimated, median across models).
What is the price difference between the Mac Studio M3 Ultra 256GB and the NVIDIA GeForce RTX 5090?
In the US the Mac Studio M3 Ultra 256GB costs $7,999 at Amazon and the NVIDIA GeForce RTX 5090 $7,489 at Amazon: the NVIDIA GeForce RTX 5090 is $510 cheaper.

More comparisons