Hardware for open-weight models
Local LLM machines compared: memory, bandwidth and measured speed
Complete machines you can order today and the DIY builds people put together instead, each checked against 41 open models for whether the 4-bit file fits in the memory its GPU can use, and how many tokens a second people have measured on the same chip or cards.
24 machines from 11 makers · 12 DIY builds · 41 models · 1044 measurements · priced ~$1,269 to $159,283
gpt-oss-120b at 4-bit, measured
| RTX PRO 6000 Blackwell Workstation Edition | 202 tok/s | n=4 |
| M5 Ultra (36-core CPU, 80-core GPU) | 111 tok/s | n=1 |
| M5 Ultra (30-core CPU, 64-core GPU) | 100 tok/s | n=3 |
| M5 Max (18-core CPU, 40-core GPU) | 90 tok/s | n=4 |
| GB10 Grace Blackwell | 53 tok/s | n=23 |
| Ryzen AI Max+ 395 | 52 tok/s | n=5 |
Medians of single-user runs with plain decoding. The order follows memory bandwidth more than compute: why bandwidth sets generation speed.
Every machine in the catalog
36 of 36 shown
CPU + RAM server
EPYC 9355P + 768 GB 12-channel DDR5 (CPU only)- Memory for models
- 760 GBassumed
- Bandwidth
- 614.4 GB/s
- Price
- ~$3,663 + unpriced parts
- motherboard, ram and cooling prices not included
- gpt-oss-20b, 4-bit
- –
Largest 4-bit model that fits: Kimi K2.6
NVIDIA workstation
Dell Pro Precision with GB300- Memory for models
- 252 GB
- Bandwidth
- 7,100 GB/s
- Price
- ~$159,283
- gpt-oss-20b, 4-bit
- ~1,678–2,143 tok/sestimate
Largest 4-bit model that fits: MiniMax M3
Apple unified memory
Apple Mac Studio (M5 Ultra, 64-core GPU)- Memory for models
- 192 GBassumed
- Bandwidth
- 1,200 GB/s
- Price
- ~$5,499–$9,499
- gpt-oss-20b, 4-bit
- 171 tok/s
Largest 4-bit model that fits: GLM-5.3 Flash
Apple unified memory
Apple Mac Studio (M5 Ultra, 80-core GPU)- Memory for models
- 192 GBassumed
- Bandwidth
- 1,200 GB/s
- Price
- ~$6,799–$10,799
- gpt-oss-20b, 4-bit
- ~200–357 tok/sestimate
Largest 4-bit model that fits: GLM-5.3 Flash
GPU desktop (prebuilt)
tiny corp tinybox red flash- Memory for models
- 184.5 GB
- Bandwidth per card
- 640 GB/s
- Price
- ~$22,000
- gpt-oss-20b, 4-bit
- ~80–117 tok/sestimate
Largest 4-bit model that fits: Hy3
AMD unified memory
Framework Desktop (Ryzen AI Max+ PRO 495)- Memory for models
- 144 GBassumed
- Bandwidth
- 273 GB/s
- Price
- ~$6,799
- gpt-oss-20b, 4-bit
- ~65–82 tok/sestimate
Largest 4-bit model that fits: MiniMax M2.7
Multi-GPU
4x Instinct MI50 32GB (used) on EPYC 7003- Memory for models
- 123.5 GB
- Bandwidth per card
- 1,024 GB/s
- Price
- ~$1,998 + unpriced parts
- GPU, motherboard, ram, cooling and case prices not included
- gpt-oss-20b, 4-bit
- ~128–187 tok/sestimate
Largest 4-bit model that fits: Qwen3.8 Flash-Next
NVIDIA GB10
NVIDIA DGX Spark- Memory for models
- 120 GBassumed
- Bandwidth
- 273 GB/s
- Price
- ~$4,699
- gpt-oss-20b, 4-bit
- 83 tok/s
Largest 4-bit model that fits: Qwen3.8 Flash-Next
NVIDIA GB10
Dell Pro Precision with GB10- Memory for models
- 120 GBassumed
- Bandwidth
- 273 GB/s
- Price
- ~$5,523
- gpt-oss-20b, 4-bit
- 83 tok/s
Largest 4-bit model that fits: Qwen3.8 Flash-Next
Other
Microsoft Surface RTX Spark Dev Box- Memory for models
- 120 GBassumed
- Bandwidth
- –
- Price
- ~$6,000
- gpt-oss-20b, 4-bit
- –
Largest 4-bit model that fits: Qwen3.8 Flash-Next
Apple unified memory
Apple Mac Studio (M5 Max, 40-core GPU)- Memory for models
- 96 GBassumed
- Bandwidth
- 614 GB/s
- Price
- ~$3,099–$5,099
- gpt-oss-20b, 4-bit
- 132 tok/s
Largest 4-bit model that fits: Nemotron 3 Super 120B-A12B
Apple unified memory
Apple MacBook Pro 14-inch (M5 Max, 40-core GPU)- Memory for models
- 96 GBassumed
- Bandwidth
- 614 GB/s
- Price
- ~$4,699–$6,699
- gpt-oss-20b, 4-bit
- 132 tok/s
Largest 4-bit model that fits: Nemotron 3 Super 120B-A12B
Apple unified memory
Apple MacBook Pro 16-inch (M5 Max, 40-core GPU)- Memory for models
- 96 GBassumed
- Bandwidth
- 614 GB/s
- Price
- ~$4,999–$6,999
- gpt-oss-20b, 4-bit
- 132 tok/s
Largest 4-bit model that fits: Nemotron 3 Super 120B-A12B
AMD unified memory
Framework Desktop (Ryzen AI Max+ 395)- Memory for models
- 96 GB
- Bandwidth
- 256 GB/s
- Price
- ~$1,959–$3,449
- gpt-oss-20b, 4-bit
- 73 tok/s
Largest 4-bit model that fits: Nemotron 3 Super 120B-A12B
AMD unified memory
GMKtec EVO-X2 (Ryzen AI Max+ 395)- Memory for models
- 96 GB
- Bandwidth
- 256 GB/s
- Price
- ~$2,600–$4,000
- gpt-oss-20b, 4-bit
- 73 tok/s
Largest 4-bit model that fits: Nemotron 3 Super 120B-A12B
AMD unified memory
Beelink GTR9 Pro (Ryzen AI Max+ 395)- Memory for models
- 96 GB
- Bandwidth
- 256 GB/s
- Price
- ~$4,699
- gpt-oss-20b, 4-bit
- 73 tok/s
Largest 4-bit model that fits: Nemotron 3 Super 120B-A12B
AMD unified memory
Minisforum MS-S1 MAX (Ryzen AI Max+ 395)- Memory for models
- 96 GB
- Bandwidth
- 256 GB/s
- Price
- ~$4,899
- gpt-oss-20b, 4-bit
- 73 tok/s
Largest 4-bit model that fits: Nemotron 3 Super 120B-A12B
AMD unified memory
Corsair AI Workstation 300 (Ryzen AI Max+ 395)- Memory for models
- 96 GB
- Bandwidth
- 256 GB/s
- Price
- ~$4,800
- gpt-oss-20b, 4-bit
- 73 tok/s
Largest 4-bit model that fits: Nemotron 3 Super 120B-A12B
AMD unified memory
HP Z2 Mini G1a (Ryzen AI Max+ PRO 395)- Memory for models
- 96 GB
- Bandwidth
- 256 GB/s
- Price
- ~$4,999
- gpt-oss-20b, 4-bit
- 73 tok/s
Largest 4-bit model that fits: Nemotron 3 Super 120B-A12B
NVIDIA workstation
Dell Pro Max Tower T2 with NVIDIA RTX PRO 6000 Blackwell- Memory for models
- 96 GB
- Bandwidth
- 1,792 GB/s
- Price
- ~$13,909
- with 8GB of system RAM; 128GB adds $3,940
- gpt-oss-20b, 4-bit
- 287 tok/s
Largest 4-bit model that fits: Nemotron 3 Super 120B-A12B
Single GPU
1x RTX PRO 6000 Blackwell- Memory for models
- 96 GB
- Bandwidth
- 1,792 GB/s
- Price
- ~$20,833
- sum of parts
- gpt-oss-20b, 4-bit
- 287 tok/s
Largest 4-bit model that fits: Nemotron 3 Super 120B-A12B
Multi-GPU
4x RTX 3090 (used) on EPYC 7003, open frame- Memory for models
- 91.5 GB
- Bandwidth per card
- 936 GB/s
- Price
- ~$2,178 + unpriced parts
- GPU, motherboard, ram, risers, case and cooling prices not included
- gpt-oss-20b, 4-bit
- ~117–171 tok/sestimate
Largest 4-bit model that fits: Nemotron 3 Super 120B-A12B
Multi-GPU
2x RTX 5090- Memory for models
- 62.5 GB
- Bandwidth per card
- 1,792 GB/s
- Price
- ~$4,109 + unpriced parts
- GPU price not included
- gpt-oss-20b, 4-bit
- ~224–327 tok/sestimate
Largest 4-bit model that fits: Llama 4 Scout 17B-16E Instruct
Apple unified memory
Apple Mac mini (M5 Pro, 16-core GPU)- Memory for models
- 48 GBassumed
- Bandwidth
- 307 GB/s
- Price
- ~$1,699–$2,699
- gpt-oss-20b, 4-bit
- 80 tok/s
Largest 4-bit model that fits: Qwen3 Coder Next
Apple unified memory
Apple Mac mini (M5 Pro, 20-core GPU)- Memory for models
- 48 GBassumed
- Bandwidth
- 307 GB/s
- Price
- ~$1,899–$2,899
- gpt-oss-20b, 4-bit
- 80 tok/s
Largest 4-bit model that fits: Qwen3 Coder Next
Multi-GPU
2x RTX 3090 (used)- Memory for models
- 46.5 GB
- Bandwidth per card
- 936 GB/s
- Price
- ~$3,639 + unpriced parts
- GPU and risers prices not included
- gpt-oss-20b, 4-bit
- ~117–171 tok/sestimate
Largest 4-bit model that fits: Llama 3.3 70B Instruct
Multi-GPU
2x Radeon RX 7900 XTX- Memory for models
- 46.5 GB
- Bandwidth per card
- 960 GB/s
- Price
- ~$5,637
- sum of parts
- gpt-oss-20b, 4-bit
- ~120–175 tok/sestimate
Largest 4-bit model that fits: Llama 3.3 70B Instruct
Multi-GPU
2x Intel Arc Pro B60- Memory for models
- 46.5 GB
- Bandwidth per card
- 456 GB/s
- Price
- ~$4,367
- sum of parts
- gpt-oss-20b, 4-bit
- ~57–83 tok/sestimate
Largest 4-bit model that fits: Llama 3.3 70B Instruct
Single GPU
1x RTX 5090- Memory for models
- 32 GB
- Bandwidth
- 1,792 GB/s
- Price
- ~$3,234 + unpriced parts
- GPU price not included
- gpt-oss-20b, 4-bit
- 290 tok/s
Largest 4-bit model that fits: Nemotron 3.5 Lightning 30B-A3B
Hybrid GPU + RAM
RTX 5090 + EPYC 9355P with 768 GB DDR5 (MoE offload)- Memory for models
- 32 GB
- Bandwidth
- 1,792 GB/s
- Price
- ~$3,763 + unpriced parts
- GPU, motherboard, ram and cooling prices not included
- gpt-oss-20b, 4-bit
- 290 tok/s
Largest 4-bit model that fits: Nemotron 3.5 Lightning 30B-A3B
Apple unified memory
Apple Mac Studio (M5 Max, 32-core GPU)- Memory for models
- 24 GBassumed
- Bandwidth
- 460 GB/s
- Price
- ~$2,499
- gpt-oss-20b, 4-bit
- 105 tok/s
Largest 4-bit model that fits: Qwen3.6 35B-A3B
Apple unified memory
Apple MacBook Pro 14-inch (M5 Max, 32-core GPU)- Memory for models
- 24 GBassumed
- Bandwidth
- 460 GB/s
- Price
- ~$4,099
- gpt-oss-20b, 4-bit
- 105 tok/s
Largest 4-bit model that fits: Qwen3.6 35B-A3B
Apple unified memory
Apple MacBook Pro 16-inch (M5 Max, 32-core GPU)- Memory for models
- 24 GBassumed
- Bandwidth
- 460 GB/s
- Price
- ~$4,399
- gpt-oss-20b, 4-bit
- 105 tok/s
Largest 4-bit model that fits: Qwen3.6 35B-A3B
AMD unified memory
Framework Desktop (Ryzen AI Max 385)- Memory for models
- 24 GBassumed
- Bandwidth
- 256 GB/s
- Price
- ~$1,269
- gpt-oss-20b, 4-bit
- ~61–77 tok/sestimate
Largest 4-bit model that fits: Qwen3.6 35B-A3B
Single GPU
1x RTX 3090 (used)- Memory for models
- 24 GB
- Bandwidth
- 936 GB/s
- Price
- ~$2,944 + unpriced parts
- used GPU price not included
- gpt-oss-20b, 4-bit
- 162 tok/s
Largest 4-bit model that fits: Qwen3.6 35B-A3B
Hybrid GPU + RAM
RTX 3090 (used) + 128 GB DDR5 on AM5 (MoE offload)- Memory for models
- 24 GB
- Bandwidth
- 936 GB/s
- Price
- ~$4,833 + unpriced parts
- used GPU price not included
- gpt-oss-20b, 4-bit
- 162 tok/s
Largest 4-bit model that fits: Qwen3.6 35B-A3B
Background reading
Why memory bandwidth sets generation speed
Each new token reads every active weight once. Compute matters for reading the prompt; bandwidth matters for writing the answer.
Quantization: what 4-bit and 8-bit files cost and save
A quant stores each weight in fewer bits. File size, and so memory and speed, fall almost in step with the bit width.
Context memory: why long conversations need more RAM
The KV cache stores every token already read. On older models it can outgrow the weights at long context.
Mixture-of-experts models: total versus active parameters
An MoE model needs memory for all of its weights but reads only the active ones per token, so it runs far faster than its size suggests.
Unified memory: how much of it the GPU can use
A 128 GB Mac or Ryzen AI Max box cannot hand all 128 GB to the GPU by default. What each platform documents, and what this site assumes.