readygpu

Mac Studio M3 Ultra 96GB vs NVIDIA DGX Spark 128GB

For running models locally, priced in Canada. At 128k context the first runs 68 models fully in memory, the second 79.

In short

The NVIDIA DGX Spark 128GB has more memory for models: 128 GB against 96 GB. At 128k context the NVIDIA DGX Spark 128GB runs 79 of the models we track fully in memory and the Mac Studio M3 Ultra 96GB 68; only the NVIDIA DGX Spark 128GB holds DeepSeek V4 Flash. On the models both run fully in memory, the Mac Studio M3 Ultra 96GB is about 3.1× as fast as the NVIDIA DGX Spark 128GB (estimated, median across models).

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

Mac Studio M3 Ultra 96GB

Not priced here

Add to build
NVIDIA DGX Spark 128GB

$11,014

Amazon

Add to build

Specifications

SpecMac Studio M3 Ultra 96GBNVIDIA DGX Spark 128GB
ChipApple M3 Ultra (80-core GPU)GB10 Grace Blackwell
Memory bus1024-bit256-bit
Memory bandwidth819 GB/s273 GB/s
FP16 compute28 TFLOPS62 TFLOPS
SoftwareMetalCUDA
Unified memory96 GB128 GB
GPU can use by default72 GB119 GB
GPU can use, tuned88 GB122 GB
Operating systemmacosdgx-os
ProcessorM3 Ultra 32-core20-core Arm (GB10)
Storage1000 GB4000 GB

What each runs

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

ModelMac Studio M3 Ultra 96GBNVIDIA DGX Spark 128GB
DeepSeek V4 Flashwon't run6tok/sUD-Q2_K_XL
GLM 5.3 Flashwon't run10tok/sUD-Q2_K_XL
Qwen 3.8 Flash Nextwon't run2tok/sQ4_K_M
Qwen 3.8 27B14tok/sQ8_05tok/sQ8_0
MiniMax M2.7won't run3tok/sUD-Q2_K_XL
Qwen 3.5 122B-A10B41tok/sQ3_K_M9tok/sQ6_K
Qwen 3.6 27B14tok/sQ8_05tok/sQ8_0
Qwen3 Coder Next49tok/sQ6_K14tok/sQ8_0
Nemotron 3 Super 120B-A12B44tok/sUD-Q3_K_M8tok/sUD-Q6_K
gpt-oss 120B40tok/sMXFP413tok/sMXFP4
Gemma 4 31B12tok/sQ8_04tok/sQ8_0
Mistral Small 4 119B44tok/sQ4_K_M11tok/sUD-Q6_K
Qwen 3.6 35B-A3B55tok/sQ8_017tok/sQ8_0
Qwen 3.5 27B14tok/sQ8_05tok/sQ8_0
Gemma 4 26B-A4B45tok/sQ8_014tok/sQ8_0
GLM 4.5 Air11tok/sQ2_K3tok/sQ6_K
Qwen3 Next 80B-A3B56tok/sQ6_K16tok/sQ8_0
Qwen 3.5 35B-A3B55tok/sQ8_017tok/sQ8_0
GLM 4.7 Flash26tok/sQ8_08tok/sQ8_0
Devstral Small 2 24B11tok/sQ8_04tok/sQ8_0
Nemotron 3.5 Lightning 30B-A3B76tok/sUD-Q6_K_XL24tok/sUD-Q6_K_XL
Kimi Linear 48B-A3B75tok/sQ8_024tok/sQ8_0
gpt-oss 20B57tok/sMXFP418tok/sMXFP4
Qwen3 Coder 30B-A3B20tok/sQ8_06tok/sQ8_0
Qwen3 32B8tok/sQ8_02tok/sQ8_0
Gemma 4 12B35tok/sQ8_011tok/sQ8_0
Qwen 3.5 9B38tok/sQ8_012tok/sQ8_0
Gemma 3 27B13tok/sQ8_04tok/sQ8_0
Mistral Small 3.2 24B11tok/sQ8_04tok/sQ8_0
DeepSeek R1 Distill Qwen 32B8tok/sQ8_02tok/sQ8_0
Llama 3.3 70Bwon't run1tok/sQ8_0
Ministral 3 14B15tok/sQ8_05tok/sQ8_0
Qwen3 14B14tok/sQ8_05tok/sQ8_0
DeepSeek R1 0528 Qwen3 8B19tok/sQ8_06tok/sQ8_0
Qwen3 8B19tok/sQ8_06tok/sQ8_0
Phi-4 14B13tok/sQ8_04tok/sQ8_0
Qwen 3.5 4B60tok/sQ8_019tok/sQ8_0
Ministral 3 8B20tok/sQ8_06tok/sQ8_0
Granite 4.2 8B17tok/sQ8_06tok/sQ8_0
Gemma 4 E4B44tok/sQ8_014tok/sQ8_0
LFM2.5 8B-A1B43tok/sQ8_014tok/sQ8_0
Llama 3.1 8B22tok/sQ6_K7tok/sQ6_K
Qwen2.5 0.5B249tok/sQ8_079tok/sQ8_0
Qwen3.5 0.8B218tok/sQ8_069tok/sQ8_0
Qwen3.5 2B174tok/sQ5_K_M55tok/sQ5_K_M
DeepSeek R1 Distill Qwen 1.5B94tok/sQ8_030tok/sQ8_0
Mellum2.1 12B A2.5B Thinking70tok/sQ8_022tok/sQ8_0
Spark X2.5 4B67tok/smlx-4bit21tok/smlx-4bit
MiniCPM5 2B64tok/sQ8_020tok/sQ8_0
FrogNano 4B 260959tok/sQ8_019tok/sQ8_0
LFM2.5 2.6B49tok/sQ8_016tok/sQ8_0
LFM2.5 VL 3B49tok/sQ8_016tok/sQ8_0
Kimi VL A3B Thinking 250645tok/sQ8_014tok/sQ8_0
Kimi VL A3B44tok/smlx-8bit14tok/smlx-8bit
GLM 4 9B 041439tok/smlx-5bit12tok/smlx-5bit
granite vision 4.1 4b37tok/sQ8_012tok/sQ8_0
granite 4.2 3b36tok/sQ8_012tok/sQ8_0
GLM 4.6V Flash35tok/sQ8_011tok/sQ8_0
DeepSeek R1 Distill Qwen 7B34tok/sQ8_011tok/sQ8_0
GLM 4.1V 9B Thinking31tok/smlx-8bit10tok/smlx-8bit
Ministral 3 3B Instruct 251230tok/sQ8_010tok/sQ8_0
LFM2.5 24B A2B25tok/sQ8_08tok/sQ8_0
DeepSeek R1 Distill Llama 8B21tok/sQ8_07tok/sQ8_0
Ministral 3 8B Reasoning 251220tok/sQ8_06tok/sQ8_0
Qwen3.8 27B OBLITERATED18tok/smlx-4bit6tok/smlx-4bit
Muse Glimmer 30B16tok/sQ8_05tok/sQ8_0
Ministral 3 14B Reasoning 251215tok/sQ8_05tok/sQ8_0
DeepSeek R1 Distill Qwen 14B13tok/sQ8_04tok/sQ8_0
Devstral Small 250711tok/sQ8_04tok/sQ8_0
Magistral Small 250611tok/sQ8_04tok/sQ8_0
Magistral Small 250711tok/sQ8_04tok/sQ8_0
Magistral Small 250911tok/sQ8_04tok/sQ8_0
granite 4.2 30b8tok/sQ8_03tok/sQ8_0
DeepSeek R1 Distill Llama 70Bwon't run1tok/sQ8_0
Devstral 2 123B Instruct 2512won't run1tok/sQ4_K_M
GLM 4.5Vwon't run3tok/sQ4_K_M
GLM 4.6Vwon't run3tok/sQ4_K_M
Kimi Dev 72Bwon't run1tok/sQ8_0
Mistral Medium 3.5 128Bwon't run1tok/sQ4_K_M
DeepSeek R1won't runwon't run
DeepSeek R1 0528won't runwon't run
DeepSeek R1 Zerowon't runwon't run
DeepSeek V3won't runwon't run
DeepSeek V3 0324won't runwon't run
DeepSeek V3.1won't runwon't run
DeepSeek V3.1 Terminuswon't runwon't run
DeepSeek V3.2won't runwon't run
DeepSeek V4 Flash 0731won't runwon't run
DeepSeek V4 Flash Vision Expwon't runwon't run
GLM 4.5won't runwon't run
GLM 4.6won't runwon't run
GLM 4.7won't runwon't run
GLM 5won'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 M2won't runwon't run
MiniMax M2.1won't runwon't run
MiniMax M2.5won't runwon't run
MiniMax M3won't runwon't run

Questions

Which is better for local AI, the Mac Studio M3 Ultra 96GB or the NVIDIA DGX Spark 128GB?
Choose the NVIDIA DGX Spark 128GB to run bigger models. Choose the Mac Studio M3 Ultra 96GB for speed.
Which can run bigger models, the Mac Studio M3 Ultra 96GB or the NVIDIA DGX Spark 128GB?
At 128k context the NVIDIA DGX Spark 128GB runs 79 of the models we track fully in memory and the Mac Studio M3 Ultra 96GB 68; only the NVIDIA DGX Spark 128GB holds DeepSeek V4 Flash.
Which is faster for local AI, the Mac Studio M3 Ultra 96GB or the NVIDIA DGX Spark 128GB?
On the models both run fully in memory, the Mac Studio M3 Ultra 96GB is about 3.1× as fast as the NVIDIA DGX Spark 128GB (estimated, median across models).

More comparisons

More hardware for local AI