Apple Mac mini (M5 Pro, 20-core GPU) vs 1x RTX 3090 (used)
The Apple Mac mini (M5 Pro, 20-core GPU) gives a model 48 GB to work in (an assumed figure), against 24 GB on the 1x RTX 3090 (used). The 1x RTX 3090 (used) has 3.0× the memory bandwidth (936 against 307 GB/s), which sets the ceiling on generation speed. The largest 4-bit model each holds at 8,192 tokens of context: Qwen3 Coder Next on the Apple Mac mini (M5 Pro, 20-core GPU), Qwen3.6 35B-A3B on the 1x RTX 3090 (used). The 1x RTX 3090 (used) has no complete price: used GPU price not included. Rows where the two differ are tinted.
Apple unified memory
Apple Mac mini (M5 Pro, 20-core GPU)
The M5 Pro Mac mini with the larger GPU, configurable to 64GB. At 64GB it runs 30B-class models at 4-bit with room left for long context.
gpt-oss-20b, 4-bit: 80 tok/s
Single GPU
1x RTX 3090 (used)
The cheapest way to 24 GB of CUDA memory: one used RTX 3090 on a current AM5 desktop. It runs 24-32B dense models at 4-bit and small MoE models entirely on the GPU.
gpt-oss-20b, 4-bit: 162 tok/s
| Specification | Apple Mac mini (M5 Pro, 20-core GPU) | 1x RTX 3090 (used) |
|---|---|---|
| Kind | Apple unified memory · Apple | Single GPU · DIY |
| Chip or GPUs | M5 Pro (18-core CPU, 20-core GPU) | RTX 3090 |
| Memory | 24 GB, 48 GB, 64 GB | 24 GB VRAM + 64 GB RAM |
| Memory for models | 48 GB (Default GPU limit)assumption | 24 GB (VRAM) |
| Other memory figure | 56 GB (Limit raised with sysctl)assumption | none |
| Memory bandwidth | 307 GB/s | 936 GB/s |
| FP16 compute | not published | 71.2 TFLOPS |
| Price | ~$1,899–$2,899 | ~$2,944 + unpriced parts (used GPU price not included) |
| Largest at 4-bit, 8k context | Qwen3 Coder Next | Qwen3.6 35B-A3B |
| Software | Metal, MLX | CUDA, Vulkan |
| Power | 155 W | 1000 W power supply |
What each one runs
Every model in the catalog that at least one of the two holds at 4-bit with 8,192 tokens of context. Copper speeds are measured medians; outlined ones are estimates. Rows where only one fits, or where the speed differs by more than a quarter, are tinted.
| Model | Apple Mac mini (M5 Pro, 20-core GPU) | 1x RTX 3090 (used) |
|---|---|---|
| Qwen3 Coder Next80B, 3B active · 45.2 GB file | 74 tok/smedian of 2 runs | does not fit |
| Llama 3.3 70B Instruct70B · 39.6 GB file | 6.9 tok/smedian of 3 runs | does not fit |
| Nemotron 3.5 Lightning 30B-A3B30B, 3B active · 23.7 GB file | 106 tok/smedian of 2 runs | does not fit |
| Qwen3.6 35B-A3B35B, 3B active · 20.8 GB file | ~53–95 tok/sestimate | ~122–178 tok/sestimate |
| Ornith 1.5 35B-A3B35B, 3B active · 20.2 GB file | ~55–98 tok/sestimate | ~125–183 tok/sestimate |
| Qwen3 Coder 30B-A3B Instruct30.5B, 3.3B active · 17.3 GB file | 95 tok/smedian of 1 run | ~116–169 tok/sestimate |
| Gemma 4 31B IT30.7B · 17.1 GB file | 13 tok/smedian of 3 runs | 35 tok/smedian of 1 run |
| Granite 4.2 30B29.3B · 16.8 GB file | ~13–16 tok/sestimate | ~25–40 tok/sestimate |
| Qwen3.8 27B27B · 16.2 GB file | 14 tok/smedian of 1 run | ~26–41 tok/sestimate |
| Gemma 4 26B-A4B IT25.2B, 3.8B active · 15.9 GB file | 71 tok/smedian of 2 runs | 119 tok/smedian of 1 run |
| Qwen3.6 27B27B · 15.7 GB file | 13 tok/smedian of 3 runs | ~27–43 tok/sestimate |
| Devstral Small 2 24B (2512)24B · 13.3 GB file | ~16–20 tok/sestimate | ~32–50 tok/sestimate |
| gpt-oss-20b21B, 3.6B active · 10.8 GB file | 80 tok/smedian of 3 runs | 162 tok/smedian of 3 runs |
| Phi-414.7B · 8.3 GB file | ~25–32 tok/sestimate | ~52–81 tok/sestimate |
| Ministral 3 14B Reasoning (2512)13.9B · 7.7 GB file | ~27–34 tok/sestimate | ~56–87 tok/sestimate |
| Gemma 4 12B IT11.95B · 6.6 GB file | ~32–39 tok/sestimate | ~64–101 tok/sestimate |
| Qwen3.5 9B9B · 5.3 GB file | 56 tok/smedian of 1 run | ~81–127 tok/sestimate |
| Granite 4.2 8B8.8B · 5.2 GB file | ~41–51 tok/sestimate | ~83–130 tok/sestimate |
| Gemma 4 E4B IT8B · 5.0 GB file | ~74–92 tok/sestimate | ~151–237 tok/sestimate |
| Qwen3 8B8.2B · 4.7 GB file | ~45–56 tok/sestimate | 115 tok/smedian of 1 run |
| Llama 3.1 8B Instruct8B · 4.6 GB file | 45 tok/smedian of 1 run | 122 tok/smedian of 2 runs |
| Qwen3.5 4B4B · 2.6 GB file | ~82–102 tok/sestimate | ~167–263 tok/sestimate |
Full pages: Apple Mac mini (M5 Pro, 20-core GPU) · 1x RTX 3090 (used) · check a different quant or context length