Apple Mac Studio (M5 Max, 40-core GPU) vs 2x RTX 3090 (used)
The Apple Mac Studio (M5 Max, 40-core GPU) gives a model 96 GB to work in (an assumed figure), against 46.5 GB on the 2x RTX 3090 (used). The 2x RTX 3090 (used) has 1.5× the memory bandwidth (936 against 614 GB/s, per card for multi-GPU builds), which sets the ceiling on generation speed. The largest 4-bit model each holds at 8,192 tokens of context: Nemotron 3 Super 120B-A12B on the Apple Mac Studio (M5 Max, 40-core GPU), Llama 3.3 70B Instruct on the 2x RTX 3090 (used). The 2x RTX 3090 (used) has no complete price: GPU and risers prices not included. Rows where the two differ are tinted.
Apple unified memory
Apple Mac Studio (M5 Max, 40-core GPU)
Mac Studio with the full M5 Max, configurable to 128GB of unified memory. At 128GB it holds a 70B dense model or a 120B-class MoE model at 4-bit inside the default GPU limit.
gpt-oss-20b, 4-bit: 132 tok/s
Multi-GPU
2x RTX 3090 (used)
Two used 3090s give 48 GB for 70B-class models at 4-bit. The 3090 is the newest GeForce card with an NVLink connector, which only matters for tensor-parallel serving.
gpt-oss-20b, 4-bit: ~117–171 tok/sestimate
| Specification | Apple Mac Studio (M5 Max, 40-core GPU) | 2x RTX 3090 (used) |
|---|---|---|
| Kind | Apple unified memory · Apple | Multi-GPU · DIY |
| Chip or GPUs | M5 Max (18-core CPU, 40-core GPU) | 2× RTX 3090 |
| Memory | 48 GB, 64 GB, 128 GB | 48 GB VRAM + 64 GB RAM |
| Memory for models | 96 GB (Default GPU limit)assumption | 46.5 GB (VRAM across all cards) |
| Other memory figure | 120 GB (Limit raised with sysctl)assumption | none |
| Memory bandwidth | 614 GB/s | 936 GB/s per card |
| FP16 compute | not published | 71.2 TFLOPS per card |
| Price | ~$3,099–$5,099 | ~$3,639 + unpriced parts (GPU and risers prices not included) |
| Largest at 4-bit, 8k context | Nemotron 3 Super 120B-A12B | Llama 3.3 70B Instruct |
| Software | Metal, MLX | CUDA, Vulkan |
| Power | 480 W | 1500 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 Studio (M5 Max, 40-core GPU) | 2x RTX 3090 (used) |
|---|---|---|
| Qwen3.8 Flash-Next125B, 6B active · 111.4 GB file | Fits with raised limitassumed limit39 tok/smedian of 2 runs | does not fit |
| Nemotron 3 Super 120B-A12B120B, 12B active · 81.0 GB file | 58 tok/smedian of 1 run | does not fit |
| Qwen3.5 122B-A10B122B, 10B active · 71.3 GB file | 58 tok/smedian of 4 runs | does not fit |
| Mistral Medium 3.5 128B128B · 69.8 GB file | ~6.0–7.5 tok/sestimate | does not fit |
| GLM-4.5-Air106B, 12B active · 68.0 GB file | 53 tok/smedian of 1 run | does not fit |
| Mistral Small 4 119B (2603)119B, 6.5B active · 67.6 GB file | 92 tok/smedian of 4 runs | does not fit |
| Llama 4 Scout 17B-16E Instruct109B, 17B active · 60.9 GB file | ~20–36 tok/sestimate | does not fit |
| gpt-oss-120b117B, 5.1B active · 58.5 GB file | 90 tok/smedian of 4 runs | does not fit |
| Qwen3 Coder Next80B, 3B active · 45.2 GB file | 98 tok/smedian of 3 runs | does not fit |
| Llama 3.3 70B Instruct70B · 39.6 GB file | 13 tok/smedian of 3 runs | 20 tok/smedian of 1 run |
| Nemotron 3.5 Lightning 30B-A3B30B, 3B active · 23.7 GB file | 142 tok/smedian of 2 runs | ~91–133 tok/sestimate |
| Qwen3.6 35B-A3B35B, 3B active · 20.8 GB file | 148 tok/smedian of 2 runs | ~122–178 tok/sestimate |
| Ornith 1.5 35B-A3B35B, 3B active · 20.2 GB file | 138 tok/smedian of 1 run | ~125–183 tok/sestimate |
| Qwen3 Coder 30B-A3B Instruct30.5B, 3.3B active · 17.3 GB file | 107 tok/smedian of 1 run | ~116–169 tok/sestimate |
| Gemma 4 31B IT30.7B · 17.1 GB file | 21 tok/smedian of 3 runs | ~25–39 tok/sestimate |
| Granite 4.2 30B29.3B · 16.8 GB file | ~25–31 tok/sestimate | ~25–40 tok/sestimate |
| Qwen3.8 27B27B · 16.2 GB file | 34 tok/smedian of 2 runs | ~26–41 tok/sestimate |
| Gemma 4 26B-A4B IT25.2B, 3.8B active · 15.9 GB file | 116 tok/smedian of 3 runs | ~91–132 tok/sestimate |
| Qwen3.6 27B27B · 15.7 GB file | 33 tok/smedian of 1 run | ~27–43 tok/sestimate |
| Devstral Small 2 24B (2512)24B · 13.3 GB file | 38 tok/smedian of 1 run | ~32–50 tok/sestimate |
| gpt-oss-20b21B, 3.6B active · 10.8 GB file | 132 tok/smedian of 4 runs | ~117–171 tok/sestimate |
| Phi-414.7B · 8.3 GB file | ~51–63 tok/sestimate | ~52–81 tok/sestimate |
| Ministral 3 14B Reasoning (2512)13.9B · 7.7 GB file | ~55–68 tok/sestimate | ~56–87 tok/sestimate |
| Gemma 4 12B IT11.95B · 6.6 GB file | 44 tok/smedian of 1 run | ~64–101 tok/sestimate |
| Qwen3.5 9B9B · 5.3 GB file | 103 tok/smedian of 1 run | ~81–127 tok/sestimate |
| Granite 4.2 8B8.8B · 5.2 GB file | ~81–101 tok/sestimate | ~83–130 tok/sestimate |
| Gemma 4 E4B IT8B · 5.0 GB file | ~148–185 tok/sestimate | ~151–237 tok/sestimate |
| Qwen3 8B8.2B · 4.7 GB file | 92 tok/smedian of 1 run | ~91–143 tok/sestimate |
| Llama 3.1 8B Instruct8B · 4.6 GB file | 90 tok/smedian of 2 runs | ~93–146 tok/sestimate |
| Qwen3.5 4B4B · 2.6 GB file | 128 tok/smedian of 1 run | ~167–263 tok/sestimate |
Full pages: Apple Mac Studio (M5 Max, 40-core GPU) · 2x RTX 3090 (used) · check a different quant or context length