readygpu

Beelink GTR9 Pro Ryzen AI Max+ 395 128GB vs Mac Studio M3 Ultra 512GB

For running models locally, priced in Germany. At 128k context the first runs 67 models fully in memory, the second 81.

In short

The Mac Studio M3 Ultra 512GB has more memory for models: 512 GB against 128 GB. At 128k context the Mac Studio M3 Ultra 512GB runs 81 of the models we track fully in memory and the Beelink GTR9 Pro Ryzen AI Max+ 395 128GB 67; only the Mac Studio M3 Ultra 512GB holds DeepSeek V4 Flash. On the models both run fully in memory, the Mac Studio M3 Ultra 512GB is about 3.8× as fast as the Beelink GTR9 Pro Ryzen AI Max+ 395 128GB (estimated, median across models).

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

Beelink GTR9 Pro Ryzen AI Max+ 395 128GB

Not priced here

Add to build
Mac Studio M3 Ultra 512GB

18.995 €

eBay, used (asking)

Add to build

Specifications

SpecBeelink GTR9 Pro Ryzen AI Max+ 395 128GBMac Studio M3 Ultra 512GB
ChipRyzen AI Max+ 395 (Radeon 8060S)Apple M3 Ultra (80-core GPU)
Memory bus256-bit1024-bit
Memory bandwidth256 GB/s819 GB/s
FP16 compute59.4 TFLOPS28 TFLOPS
SoftwareROCmMetal
Unified memory128 GB512 GB
GPU can use by default96 GB384 GB
GPU can use, tuned120 GB504 GB
Operating systemwindowsmacos
ProcessorRyzen AI Max+ 395M3 Ultra 32-core
Storage2000 GB2000 GB

What each runs

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

ModelBeelink GTR9 Pro Ryzen AI Max+ 395 128GBMac Studio M3 Ultra 512GB
DeepSeek V4 Flashwon't run14tok/sMXFP4
Qwen 3.8 Flash Next3tok/sUD-Q3_K_XL5tok/sQ8_0
GLM 5.3 Flashwon't run11tok/sQ8_0
Qwen 3.8 27B4tok/sQ8_014tok/sQ8_0
MiniMax M2.7won't run7tok/sQ8_0
Qwen 3.5 122B-A10B8tok/sQ5_K_M23tok/sQ8_0
Qwen 3.6 27B4tok/sQ8_014tok/sQ8_0
Qwen3 Coder Next11tok/sQ8_043tok/sQ8_0
gpt-oss 120B11tok/sMXFP440tok/sMXFP4
Nemotron 3 Super 120B-A12B8tok/sQ5_K_M23tok/sQ8_0
Gemma 4 31B3tok/sQ8_012tok/sQ8_0
Mistral Small 4 119B10tok/sQ6_K30tok/sQ8_0
Qwen 3.6 35B-A3B15tok/sQ8_055tok/sQ8_0
Qwen 3.5 27B4tok/sQ8_014tok/sQ8_0
Gemma 4 26B-A4B12tok/sQ8_045tok/sQ8_0
GLM 4.5 Air2tok/sQ4_K_M8tok/sQ8_0
Qwen 3.5 35B-A3B15tok/sQ8_055tok/sQ8_0
Qwen3 Next 80B-A3B13tok/sQ8_050tok/sQ8_0
GLM 4.7 Flash7tok/sQ8_026tok/sQ8_0
Devstral Small 2 24B3tok/sQ8_011tok/sQ8_0
Nemotron 3.5 Lightning 30B-A3B20tok/sUD-Q6_K_XL76tok/sUD-Q6_K_XL
Kimi Linear 48B-A3B20tok/sQ8_075tok/sQ8_0
gpt-oss 20B15tok/sMXFP457tok/sMXFP4
Qwen3 Coder 30B-A3B5tok/sQ8_020tok/sQ8_0
Qwen3 32B2tok/sQ8_08tok/sQ8_0
Gemma 4 12B9tok/sQ8_035tok/sQ8_0
Qwen 3.5 9B10tok/sQ8_038tok/sQ8_0
Gemma 3 27B4tok/sQ8_013tok/sQ8_0
Mistral Small 3.2 24B3tok/sQ8_011tok/sQ8_0
DeepSeek R1 Distill Qwen 32B2tok/sQ8_08tok/sQ8_0
Llama 3.3 70B2tok/sQ5_K_M5tok/sQ8_0
Ministral 3 14B4tok/sQ8_015tok/sQ8_0
Qwen3 14B4tok/sQ8_014tok/sQ8_0
DeepSeek R1 0528 Qwen3 8B5tok/sQ8_019tok/sQ8_0
Qwen3 8B5tok/sQ8_019tok/sQ8_0
Phi-4 14B3tok/sQ8_013tok/sQ8_0
Qwen 3.5 4B16tok/sQ8_060tok/sQ8_0
Ministral 3 8B5tok/sQ8_020tok/sQ8_0
Granite 4.2 8B5tok/sQ8_017tok/sQ8_0
Gemma 4 E4B12tok/sQ8_044tok/sQ8_0
LFM2.5 8B-A1B11tok/sQ8_043tok/sQ8_0
Llama 3.1 8B6tok/sQ6_K22tok/sQ6_K
Qwen2.5 0.5B66tok/sQ8_0249tok/sQ8_0
Qwen3.5 0.8B58tok/sQ8_0218tok/sQ8_0
Qwen3.5 2B46tok/sQ5_K_M174tok/sQ5_K_M
DeepSeek R1 Distill Qwen 1.5B25tok/sQ8_094tok/sQ8_0
Mellum2.1 12B A2.5B Thinking18tok/sQ8_070tok/sQ8_0
MiniCPM5 2B17tok/sQ8_064tok/sQ8_0
LFM2.5 2.6B13tok/sQ8_049tok/sQ8_0
LFM2.5 VL 3B13tok/sQ8_049tok/sQ8_0
Kimi VL A3B Thinking 250612tok/sQ8_045tok/sQ8_0
Kimi VL A3B12tok/smlx-8bit44tok/smlx-8bit
GLM 4 9B 041410tok/smlx-5bit39tok/smlx-5bit
granite vision 4.1 4b10tok/sQ8_037tok/sQ8_0
granite 4.2 3b10tok/sQ8_036tok/sQ8_0
GLM 4.6V Flash9tok/sQ8_035tok/sQ8_0
DeepSeek R1 Distill Qwen 7B9tok/sQ8_034tok/sQ8_0
GLM 4.1V 9B Thinking8tok/smlx-8bit31tok/smlx-8bit
LFM2.5 24B A2B7tok/sQ8_025tok/sQ8_0
DeepSeek R1 Distill Llama 8B6tok/sQ8_021tok/sQ8_0
Ministral 3 8B Reasoning 25125tok/sQ8_020tok/sQ8_0
DeepSeek R1 Distill Qwen 14B3tok/sQ8_013tok/sQ8_0
Devstral Small 25073tok/sQ8_011tok/sQ8_0
Magistral Small 25063tok/sQ8_011tok/sQ8_0
Magistral Small 25073tok/sQ8_011tok/sQ8_0
Magistral Small 25093tok/sQ8_011tok/sQ8_0
GLM 4.5V3tok/sQ4_K_M8tok/sQ8_0
GLM 4.6V3tok/sQ4_K_M8tok/sQ8_0
granite 4.2 30b2tok/sQ8_08tok/sQ8_0
DeepSeek R1 Distill Llama 70B2tok/sQ5_K_M5tok/sQ8_0
DeepSeek V4 Flash 0731won't run11tok/sMXFP4
DeepSeek V4 Flash Vision Expwon't run11tok/sMXFP4
Devstral 2 123B Instruct 2512won't run3tok/sQ8_0
GLM 4.6won't run4tok/sQ6_K
GLM 4.7won't run4tok/sQ6_K
Kimi Dev 72Bwon't run4tok/sQ8_0
MiniMax M2won't run7tok/sQ8_0
MiniMax M2.1won't run7tok/sQ8_0
MiniMax M2.5won't run7tok/sQ8_0
MiniMax M3won't run9tok/sUD-Q6_K
Mistral Medium 3.5 128Bwon't run3tok/sQ8_0
DeepSeek R1won't runwon't run
DeepSeek R1 0528won't runwon't run
DeepSeek R1 Zerowon'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
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

Questions

Which is better for local AI, the Beelink GTR9 Pro Ryzen AI Max+ 395 128GB or the Mac Studio M3 Ultra 512GB?
Choose the Mac Studio M3 Ultra 512GB to run bigger models. Choose the Mac Studio M3 Ultra 512GB for speed.
Which can run bigger models, the Beelink GTR9 Pro Ryzen AI Max+ 395 128GB or the Mac Studio M3 Ultra 512GB?
At 128k context the Mac Studio M3 Ultra 512GB runs 81 of the models we track fully in memory and the Beelink GTR9 Pro Ryzen AI Max+ 395 128GB 67; only the Mac Studio M3 Ultra 512GB holds DeepSeek V4 Flash.
Which is faster for local AI, the Beelink GTR9 Pro Ryzen AI Max+ 395 128GB or the Mac Studio M3 Ultra 512GB?
On the models both run fully in memory, the Mac Studio M3 Ultra 512GB is about 3.8× as fast as the Beelink GTR9 Pro Ryzen AI Max+ 395 128GB (estimated, median across models).

More comparisons