readygpu

Framework Desktop Ryzen AI Max+ 395 64GB vs NVIDIA DGX Spark 128GB

For running models locally, priced in Germany. At 64k context the first runs 65 models fully in memory, the second 79.

In short

The NVIDIA DGX Spark 128GB has more memory for models: 128 GB against 64 GB. At 64k context the NVIDIA DGX Spark 128GB runs 79 of the models we track fully in memory and the Framework Desktop Ryzen AI Max+ 395 64GB 65; only the NVIDIA DGX Spark 128GB holds DeepSeek V4 Flash. On the models both run fully in memory, the NVIDIA DGX Spark 128GB is about 1.2× as fast as the Framework Desktop Ryzen AI Max+ 395 64GB (estimated, median across models).

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

Framework Desktop Ryzen AI Max+ 395 64GB

Not priced here

Add to build
NVIDIA DGX Spark 128GB

9.029 €

eBay

Add to build

Specifications

SpecFramework Desktop Ryzen AI Max+ 395 64GBNVIDIA DGX Spark 128GB
ChipRyzen AI Max+ 395 (Radeon 8060S)GB10 Grace Blackwell
Memory bus256-bit256-bit
Memory bandwidth256 GB/s273 GB/s
FP16 compute59.4 TFLOPS62 TFLOPS
SoftwareROCmCUDA
Unified memory64 GB128 GB
GPU can use by default48 GB119 GB
GPU can use, tuned56 GB122 GB
Operating systemlinuxdgx-os
ProcessorRyzen AI Max+ 39520-core Arm (GB10)
Storage1000 GB4000 GB

What each runs

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

ModelFramework Desktop Ryzen AI Max+ 395 64GBNVIDIA DGX Spark 128GB
DeepSeek V4 Flashwon't run8tok/sUD-Q2_K_XL
GLM 5.3 Flashwon't run11tok/sUD-Q2_K_XL
Qwen 3.8 Flash Nextwon't run2tok/sQ4_K_M
Qwen 3.8 27B4tok/sQ8_05tok/sQ8_0
MiniMax M2.7won't run5tok/sUD-Q3_K_M
Qwen 3.5 122B-A10B17tok/sUD-Q2_K_XL10tok/sQ6_K
Qwen 3.6 27B4tok/sQ8_05tok/sQ8_0
Qwen3 Coder Next21tok/sQ4_K_M18tok/sQ8_0
Gemma 4 31B4tok/sQ8_04tok/sQ8_0
Nemotron 3 Super 120B-A12Bwon't run9tok/sUD-Q6_K
gpt-oss 120Bwon't run18tok/sMXFP4
Mistral Small 4 119B22tok/sUD-Q2_K_XL14tok/sUD-Q6_K
Qwen 3.6 35B-A3B19tok/sQ8_023tok/sQ8_0
Qwen 3.5 27B4tok/sQ8_05tok/sQ8_0
Gemma 4 26B-A4B15tok/sQ8_018tok/sQ8_0
GLM 4.5 Airwon't run4tok/sQ6_K
Qwen3 Next 80B-A3B24tok/sQ4_K_M21tok/sQ8_0
Qwen 3.5 35B-A3B19tok/sQ8_023tok/sQ8_0
GLM 4.7 Flash10tok/sQ8_012tok/sQ8_0
Devstral Small 2 24B4tok/sQ8_05tok/sQ8_0
Nemotron 3.5 Lightning 30B-A3B22tok/sUD-Q6_K_XL27tok/sUD-Q6_K_XL
Kimi Linear 48B-A3B28tok/sQ6_K27tok/sQ8_0
gpt-oss 20B21tok/sMXFP425tok/sMXFP4
Qwen3 Coder 30B-A3B9tok/sQ8_011tok/sQ8_0
Qwen3 32B3tok/sQ6_K3tok/sQ8_0
Gemma 4 12B10tok/sQ8_012tok/sQ8_0
Qwen 3.5 9B12tok/sQ8_014tok/sQ8_0
Gemma 3 27B4tok/sQ8_05tok/sQ8_0
Mistral Small 3.2 24B4tok/sQ8_05tok/sQ8_0
DeepSeek R1 Distill Qwen 32B3tok/sQ6_K3tok/sQ8_0
Llama 3.3 70Bwon't run2tok/sQ8_0
Ministral 3 14B6tok/sQ8_07tok/sQ8_0
Qwen3 14B5tok/sQ8_06tok/sQ8_0
DeepSeek R1 0528 Qwen3 8B8tok/sQ8_09tok/sQ8_0
Qwen3 8B8tok/sQ8_09tok/sQ8_0
Phi-4 14B5tok/sQ8_06tok/sQ8_0
Qwen 3.5 4B21tok/sQ8_025tok/sQ8_0
Ministral 3 8B8tok/sQ8_09tok/sQ8_0
Granite 4.2 8B7tok/sQ8_08tok/sQ8_0
Gemma 4 E4B14tok/sQ8_017tok/sQ8_0
LFM2.5 8B-A1B20tok/sQ8_024tok/sQ8_0
Llama 3.1 8B9tok/sQ6_K11tok/sQ6_K
Qwen2.5 0.5B105tok/sQ8_0127tok/sQ8_0
Qwen3.5 0.8B86tok/sQ8_0104tok/sQ8_0
Qwen3.5 2B62tok/sQ5_K_M75tok/sQ5_K_M
DeepSeek R1 Distill Qwen 1.5B37tok/sQ8_045tok/sQ8_0
MiniCPM5 2B26tok/sQ8_031tok/sQ8_0
Spark X2.5 4B26tok/smlx-4bit31tok/smlx-4bit
Mellum2.1 12B A2.5B Thinking23tok/sQ8_028tok/sQ8_0
FrogNano 4B 260921tok/sQ8_025tok/sQ8_0
LFM2.5 2.6B20tok/sQ8_025tok/sQ8_0
LFM2.5 VL 3B20tok/sQ8_025tok/sQ8_0
Kimi VL A3B Thinking 250616tok/sQ8_020tok/sQ8_0
Kimi VL A3B16tok/smlx-8bit19tok/smlx-8bit
granite vision 4.1 4b16tok/sQ8_019tok/sQ8_0
granite 4.2 3b15tok/sQ8_018tok/sQ8_0
Ministral 3 3B Instruct 251213tok/sQ8_016tok/sQ8_0
GLM 4 9B 041413tok/smlx-5bit16tok/smlx-5bit
DeepSeek R1 Distill Qwen 7B12tok/sQ8_014tok/sQ8_0
LFM2.5 24B A2B11tok/sQ8_014tok/sQ8_0
GLM 4.6V Flash11tok/sQ8_013tok/sQ8_0
GLM 4.1V 9B Thinking10tok/smlx-8bit12tok/smlx-8bit
DeepSeek R1 Distill Llama 8B8tok/sQ8_010tok/sQ8_0
Ministral 3 8B Reasoning 25128tok/sQ8_09tok/sQ8_0
Ministral 3 14B Reasoning 25126tok/sQ8_07tok/sQ8_0
Qwen3.8 27B OBLITERATED6tok/smlx-4bit7tok/smlx-4bit
DeepSeek R1 Distill Qwen 14B5tok/sQ8_06tok/sQ8_0
Muse Glimmer 30B4tok/sQ8_05tok/sQ8_0
Devstral Small 25074tok/sQ8_05tok/sQ8_0
Magistral Small 25064tok/sQ8_05tok/sQ8_0
Magistral Small 25074tok/sQ8_05tok/sQ8_0
Magistral Small 25094tok/sQ8_05tok/sQ8_0
granite 4.2 30b3tok/sQ8_04tok/sQ8_0
DeepSeek R1 Distill Llama 70Bwon't run2tok/sQ8_0
Devstral 2 123B Instruct 2512won't run1tok/sQ6_K
GLM 4.5Vwon't run4tok/sQ8_0
GLM 4.6Vwon't run4tok/sQ8_0
Kimi Dev 72Bwon't run2tok/sQ8_0
Mistral Medium 3.5 128Bwon't run1tok/sQ6_K
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 Framework Desktop Ryzen AI Max+ 395 64GB or the NVIDIA DGX Spark 128GB?
Choose the NVIDIA DGX Spark 128GB to run bigger models. Choose the NVIDIA DGX Spark 128GB for speed.
Which can run bigger models, the Framework Desktop Ryzen AI Max+ 395 64GB or the NVIDIA DGX Spark 128GB?
At 64k context the NVIDIA DGX Spark 128GB runs 79 of the models we track fully in memory and the Framework Desktop Ryzen AI Max+ 395 64GB 65; only the NVIDIA DGX Spark 128GB holds DeepSeek V4 Flash.
Which is faster for local AI, the Framework Desktop Ryzen AI Max+ 395 64GB or the NVIDIA DGX Spark 128GB?
On the models both run fully in memory, the NVIDIA DGX Spark 128GB is about 1.2× as fast as the Framework Desktop Ryzen AI Max+ 395 64GB (estimated, median across models).

More comparisons

More hardware for local AI