Mac Studio M3 Ultra 96GB vs NVIDIA GeForce RTX 4090

For running models locally, priced in the UK. At 32k context the first runs 38 models fully in memory, the second 26.

In short

The Mac Studio M3 Ultra 96GB has more memory for models: 96 GB against 24 GB. At 32k context the Mac Studio M3 Ultra 96GB runs 38 of the models we track fully in memory and the NVIDIA GeForce RTX 4090 26; only the Mac Studio M3 Ultra 96GB holds Qwen 3.5 122B-A10B. On the models both run fully in memory, the NVIDIA GeForce RTX 4090 is about 1.2× as fast as the Mac Studio M3 Ultra 96GB (estimated, median across models). In the UK the Mac Studio M3 Ultra 96GB costs £5,499 at eBay and the NVIDIA GeForce RTX 4090 £6,015 at Amazon: the Mac Studio M3 Ultra 96GB is £516 cheaper.

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

Specifications

SpecMac Studio M3 Ultra 96GBNVIDIA GeForce RTX 4090
ChipApple M3 Ultra (80-core GPU)–
Memory bus1024-bit384-bit
Memory bandwidth819 GB/s1,008 GB/s
FP16 compute28 TFLOPS165 TFLOPS
SoftwareMetalCUDA
Unified memory96 GB–
GPU can use by default72 GB–
GPU can use, tuned88 GB–
Operating systemmacos–
ProcessorM3 Ultra 32-core–
Storage1000 GB–
GPU–GeForce RTX 4090
Memory–24 GB GDDR6X
Board power–450 W
PCIe generation–Gen 4
PCIe lanes–x16

What each runs

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

ModelMac Studio M3 Ultra 96GBNVIDIA GeForce RTX 4090
Qwen 3.8 27B17tok/sQ8_028tok/sUD-Q5_K_M
Qwen 3.5 122B-A10B58tok/sQ3_K_Moffload
Qwen 3.6 27B17tok/sQ8_029tok/sQ5_K_M
Qwen3 Coder Next78tok/sQ6_Koffload
gpt-oss 120B74tok/sMXFP4won't run
Nemotron 3 Super 120B-A12B50tok/sUD-Q3_K_Mwon't run
Gemma 4 31B15tok/sQ8_029tok/sQ4_K_M
Mistral Small 4 119B62tok/sUD-Q4_K_Moffload
Qwen 3.6 35B-A3B84tok/sQ8_0145tok/sUD-Q4_K_M
Qwen 3.5 27B17tok/sQ8_029tok/sQ5_K_M
Gemma 4 26B-A4B63tok/sQ8_084tok/sUD-Q6_K
GLM 4.5 Air21tok/sQ3_K_Moffload
Qwen3 Next 80B-A3B97tok/sQ6_Koffload
Qwen 3.5 35B-A3B84tok/sQ8_0146tok/sQ4_K_M
GLM 4.7 Flash45tok/sQ8_070tok/sQ5_K_M
Devstral Small 2 24B18tok/sQ8_028tok/sQ5_K_M
Nemotron 3.5 Lightning 30B-A3B89tok/sUD-Q6_K_XLoffload
gpt-oss 20B99tok/sMXFP4116tok/sMXFP4
Kimi Linear 48B-A3B91tok/sQ8_0offload
Qwen3 Coder 30B-A3B50tok/sQ8_074tok/sQ4_K_M
Qwen3 32B12tok/sQ8_0offload
Gemma 4 12B39tok/sQ8_046tok/sQ8_0
Qwen 3.5 9B50tok/sQ8_059tok/sQ8_0
Mistral Small 3.2 24B18tok/sQ8_028tok/sQ5_K_M
Gemma 3 27B17tok/sQ8_028tok/sQ5_K_M
DeepSeek R1 Distill Qwen 32B12tok/sQ8_0offload
Llama 3.3 70B9tok/sQ5_K_Mwon't run
Ministral 3 14B27tok/sQ8_032tok/sQ8_0
Qwen3 14B25tok/sQ8_030tok/sQ8_0
DeepSeek R1 0528 Qwen3 8B39tok/sQ8_046tok/sQ8_0
Qwen3 8B39tok/sQ8_046tok/sQ8_0
Phi-4 14B24tok/sQ8_028tok/sQ8_0
Qwen 3.5 4B95tok/sQ8_0112tok/sQ8_0
Ministral 3 8B39tok/sQ8_046tok/sQ8_0
Granite 4.2 8B36tok/sQ8_043tok/sQ8_0
Gemma 4 E4B58tok/sQ8_068tok/sQ8_0
LFM2.5 8B-A1B119tok/sQ8_0140tok/sQ8_0
Llama 3.1 8B49tok/sQ6_K57tok/sQ6_K
DeepSeek V4 Flashwon't runwon't run
GLM 5.3 Flashwon't runwon't run
Qwen 3.8 Flash Nextwon't runwon't run
MiniMax M2.7won't runwon't run

Questions

Which is better for local AI, the Mac Studio M3 Ultra 96GB or the NVIDIA GeForce RTX 4090?
Choose the Mac Studio M3 Ultra 96GB to run bigger models. Choose the NVIDIA GeForce RTX 4090 for speed. Choose the Mac Studio M3 Ultra 96GB to spend less.
Which can run bigger models, the Mac Studio M3 Ultra 96GB or the NVIDIA GeForce RTX 4090?
At 32k context the Mac Studio M3 Ultra 96GB runs 38 of the models we track fully in memory and the NVIDIA GeForce RTX 4090 26; only the Mac Studio M3 Ultra 96GB holds Qwen 3.5 122B-A10B.
Which is faster for local AI, the Mac Studio M3 Ultra 96GB or the NVIDIA GeForce RTX 4090?
On the models both run fully in memory, the NVIDIA GeForce RTX 4090 is about 1.2× as fast as the Mac Studio M3 Ultra 96GB (estimated, median across models).
What is the price difference between the Mac Studio M3 Ultra 96GB and the NVIDIA GeForce RTX 4090?
In the UK the Mac Studio M3 Ultra 96GB costs £5,499 at eBay and the NVIDIA GeForce RTX 4090 £6,015 at Amazon: the Mac Studio M3 Ultra 96GB is £516 cheaper.

More comparisons