Machines

24 machines you can order and 12 you build

Grouped by what does the work. The last column is gpt-oss-20b at 4-bit: copper figures are measured medians, outlined ones are estimates from bandwidth.

Apple unified memory

MachineMemory for modelsBandwidthPriceLargest at 4-bitgpt-oss-20b
Apple Mac Studio (M5 Max, 32-core GPU)Apple unified memory24 GBassumed460 GB/s~$2,499Qwen3.6 35B-A3B105 tok/s
Apple Mac Studio (M5 Max, 40-core GPU)Apple unified memory96 GBassumed614 GB/s~$3,099–$5,099Nemotron 3 Super 120B-A12B132 tok/s
Apple Mac Studio (M5 Ultra, 64-core GPU)Apple unified memory192 GBassumed1,200 GB/s~$5,499–$9,499GLM-5.3 Flash171 tok/s
Apple Mac Studio (M5 Ultra, 80-core GPU)Apple unified memory192 GBassumed1,200 GB/s~$6,799–$10,799GLM-5.3 Flash~200–357 tok/sestimate
Apple Mac mini (M5 Pro, 16-core GPU)Apple unified memory48 GBassumed307 GB/s~$1,699–$2,699Qwen3 Coder Next80 tok/s
Apple Mac mini (M5 Pro, 20-core GPU)Apple unified memory48 GBassumed307 GB/s~$1,899–$2,899Qwen3 Coder Next80 tok/s
Apple MacBook Pro 14-inch (M5 Max, 32-core GPU)Apple unified memory24 GBassumed460 GB/s~$4,099Qwen3.6 35B-A3B105 tok/s
Apple MacBook Pro 14-inch (M5 Max, 40-core GPU)Apple unified memory96 GBassumed614 GB/s~$4,699–$6,699Nemotron 3 Super 120B-A12B132 tok/s
Apple MacBook Pro 16-inch (M5 Max, 32-core GPU)Apple unified memory24 GBassumed460 GB/s~$4,399Qwen3.6 35B-A3B105 tok/s
Apple MacBook Pro 16-inch (M5 Max, 40-core GPU)Apple unified memory96 GBassumed614 GB/s~$4,999–$6,999Nemotron 3 Super 120B-A12B132 tok/s

AMD unified memory

MachineMemory for modelsBandwidthPriceLargest at 4-bitgpt-oss-20b
Framework Desktop (Ryzen AI Max+ 395)AMD unified memory96 GB256 GB/s~$1,959–$3,449Nemotron 3 Super 120B-A12B73 tok/s
Framework Desktop (Ryzen AI Max 385)AMD unified memory24 GBassumed256 GB/s~$1,269Qwen3.6 35B-A3B~61–77 tok/sestimate
Framework Desktop (Ryzen AI Max+ PRO 495)AMD unified memory144 GBassumed273 GB/s~$6,799MiniMax M2.7~65–82 tok/sestimate
GMKtec EVO-X2 (Ryzen AI Max+ 395)AMD unified memory96 GB256 GB/s~$2,600–$4,000Nemotron 3 Super 120B-A12B73 tok/s
Beelink GTR9 Pro (Ryzen AI Max+ 395)AMD unified memory96 GB256 GB/s~$4,699Nemotron 3 Super 120B-A12B73 tok/s
Minisforum MS-S1 MAX (Ryzen AI Max+ 395)AMD unified memory96 GB256 GB/s~$4,899Nemotron 3 Super 120B-A12B73 tok/s
Corsair AI Workstation 300 (Ryzen AI Max+ 395)AMD unified memory96 GB256 GB/s~$4,800Nemotron 3 Super 120B-A12B73 tok/s
HP Z2 Mini G1a (Ryzen AI Max+ PRO 395)AMD unified memory96 GB256 GB/s~$4,999Nemotron 3 Super 120B-A12B73 tok/s

NVIDIA GB10

MachineMemory for modelsBandwidthPriceLargest at 4-bitgpt-oss-20b
NVIDIA DGX SparkNVIDIA GB10120 GBassumed273 GB/s~$4,699Qwen3.8 Flash-Next83 tok/s
Dell Pro Precision with GB10NVIDIA GB10120 GBassumed273 GB/s~$5,523Qwen3.8 Flash-Next83 tok/s

NVIDIA workstation

MachineMemory for modelsBandwidthPriceLargest at 4-bitgpt-oss-20b
Dell Pro Precision with GB300NVIDIA workstation252 GB7,100 GB/s~$159,283MiniMax M3~1,678–2,143 tok/sestimate
Dell Pro Max Tower T2 with NVIDIA RTX PRO 6000 BlackwellNVIDIA workstation96 GB1,792 GB/s~$13,909with 8GB of system RAM; 128GB adds $3,940Nemotron 3 Super 120B-A12B287 tok/s

GPU desktop (prebuilt)

MachineMemory for modelsBandwidthPriceLargest at 4-bitgpt-oss-20b
tiny corp tinybox red flashGPU desktop (prebuilt)184.5 GB640 GB/sper card~$22,000Hy3~80–117 tok/sestimate

Other

MachineMemory for modelsBandwidthPriceLargest at 4-bitgpt-oss-20b
Microsoft Surface RTX Spark Dev BoxOther120 GBassumed–~$6,000Qwen3.8 Flash-Next–

Single GPU

MachineMemory for modelsBandwidthPriceLargest at 4-bitgpt-oss-20b
1x RTX 3090 (used)Single GPU24 GB936 GB/s~$2,944 + unpriced partsused GPU price not includedQwen3.6 35B-A3B162 tok/s
1x RTX 5090Single GPU32 GB1,792 GB/s~$3,234 + unpriced partsGPU price not includedNemotron 3.5 Lightning 30B-A3B290 tok/s
1x RTX PRO 6000 BlackwellSingle GPU96 GB1,792 GB/s~$20,833sum of partsNemotron 3 Super 120B-A12B287 tok/s

Multi-GPU

MachineMemory for modelsBandwidthPriceLargest at 4-bitgpt-oss-20b
2x RTX 3090 (used)Multi-GPU46.5 GB936 GB/sper card~$3,639 + unpriced partsGPU and risers prices not includedLlama 3.3 70B Instruct~117–171 tok/sestimate
4x RTX 3090 (used) on EPYC 7003, open frameMulti-GPU91.5 GB936 GB/sper card~$2,178 + unpriced partsGPU, motherboard, ram, risers, case and cooling prices not includedNemotron 3 Super 120B-A12B~117–171 tok/sestimate
2x RTX 5090Multi-GPU62.5 GB1,792 GB/sper card~$4,109 + unpriced partsGPU price not includedLlama 4 Scout 17B-16E Instruct~224–327 tok/sestimate
2x Radeon RX 7900 XTXMulti-GPU46.5 GB960 GB/sper card~$5,637sum of partsLlama 3.3 70B Instruct~120–175 tok/sestimate
4x Instinct MI50 32GB (used) on EPYC 7003Multi-GPU123.5 GB1,024 GB/sper card~$1,998 + unpriced partsGPU, motherboard, ram, cooling and case prices not includedQwen3.8 Flash-Next~128–187 tok/sestimate
2x Intel Arc Pro B60Multi-GPU46.5 GB456 GB/sper card~$4,367sum of partsLlama 3.3 70B Instruct~57–83 tok/sestimate

Hybrid GPU + RAM

MachineMemory for modelsBandwidthPriceLargest at 4-bitgpt-oss-20b
RTX 5090 + EPYC 9355P with 768 GB DDR5 (MoE offload)Hybrid GPU + RAM32 GB1,792 GB/s~$3,763 + unpriced partsGPU, motherboard, ram and cooling prices not includedNemotron 3.5 Lightning 30B-A3B290 tok/s
RTX 3090 (used) + 128 GB DDR5 on AM5 (MoE offload)Hybrid GPU + RAM24 GB936 GB/s~$4,833 + unpriced partsused GPU price not includedQwen3.6 35B-A3B162 tok/s

CPU + RAM server

MachineMemory for modelsBandwidthPriceLargest at 4-bitgpt-oss-20b
EPYC 9355P + 768 GB 12-channel DDR5 (CPU only)CPU + RAM server760 GBassumed614.4 GB/s~$3,663 + unpriced partsmotherboard, ram and cooling prices not includedKimi K2.6–