Apple Mac Studio (M5 Ultra, 80-core GPU) vs 4x RTX 3090 (used) on EPYC 7003, open frame
The Apple Mac Studio (M5 Ultra, 80-core GPU) gives a model 192 GB to work in (an assumed figure), against 91.5 GB on the 4x RTX 3090 (used) on EPYC 7003, open frame. The Apple Mac Studio (M5 Ultra, 80-core GPU) has 1.3× the memory bandwidth (1,200 against 936 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: GLM-5.3 Flash on the Apple Mac Studio (M5 Ultra, 80-core GPU), Nemotron 3 Super 120B-A12B on the 4x RTX 3090 (used) on EPYC 7003, open frame. The 4x RTX 3090 (used) on EPYC 7003, open frame has no complete price: GPU, motherboard, ram, risers, case and cooling prices not included. Rows where the two differ are tinted.
Apple unified memory
Apple Mac Studio (M5 Ultra, 80-core GPU)
The full M5 Ultra Mac Studio. Apple lists a 512GB option for this chip as coming in late October 2026; until it ships, 256GB is the largest orderable memory.
gpt-oss-20b, 4-bit: ~200–357 tok/sestimate
Multi-GPU
4x RTX 3090 (used) on EPYC 7003, open frame
The classic 96 GB home rig: four used 3090s on risers around a single-socket EPYC board with a full x16 link for every card. It runs 70B dense models at 8-bit or ~120B MoE models in VRAM.
gpt-oss-20b, 4-bit: ~117–171 tok/sestimate
| Specification | Apple Mac Studio (M5 Ultra, 80-core GPU) | 4x RTX 3090 (used) on EPYC 7003, open frame |
|---|---|---|
| Kind | Apple unified memory · Apple | Multi-GPU · DIY |
| Chip or GPUs | M5 Ultra (36-core CPU, 80-core GPU) | 4× RTX 3090 |
| Memory | 96 GB, 256 GB | 96 GB VRAM + 256 GB RAM |
| Memory for models | 192 GB (Default GPU limit)assumption | 91.5 GB (VRAM across all cards) |
| Other memory figure | 248 GB (Limit raised with sysctl)assumption | none |
| Memory bandwidth | 1,200 GB/s | 936 GB/s per card |
| FP16 compute | not published | 71.2 TFLOPS per card |
| Price | ~$6,799–$10,799 | ~$2,178 + unpriced parts (GPU, motherboard, ram, risers, case and cooling prices not included) |
| Largest at 4-bit, 8k context | GLM-5.3 Flash | Nemotron 3 Super 120B-A12B |
| Software | Metal, MLX | CUDA, Vulkan |
| Power | 480 W | 3000 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 Ultra, 80-core GPU) | 4x RTX 3090 (used) on EPYC 7003, open frame |
|---|---|---|
| MiniMax M3428B, 23B active · 243.3 GB file | Fits with raised limitassumed limit~28–51 tok/sestimate | does not fit |
| Qwen3.5 397B-A17B397B, 17B active · 227.3 GB file | Fits with raised limitassumed limit~38–68 tok/sestimate | does not fit |
| GLM-5.3 Flash320B, 18B active · 186.0 GB file | ~36–63 tok/sestimate | does not fit |
| Hy3295B, 21B active · 169.6 GB file | ~31–55 tok/sestimate | does not fit |
| MiniMax M2.7228.7B · 129.3 GB file | – | does not fit |
| Qwen3.8 Flash-Next125B, 6B active · 111.4 GB file | ~70–124 tok/sestimate | does not fit |
| Nemotron 3 Super 120B-A12B120B, 12B active · 81.0 GB file | ~46–82 tok/sestimate | ~27–39 tok/sestimate |
| Qwen3.5 122B-A10B122B, 10B active · 71.3 GB file | ~64–113 tok/sestimate | ~37–54 tok/sestimate |
| Mistral Medium 3.5 128B128B · 69.8 GB file | ~12–15 tok/sestimate | ~6.1–9.6 tok/sestimate |
| GLM-4.5-Air106B, 12B active · 68.0 GB file | ~48–86 tok/sestimate | ~28–41 tok/sestimate |
| Mistral Small 4 119B (2603)119B, 6.5B active · 67.6 GB file | ~101–179 tok/sestimate | ~59–86 tok/sestimate |
| Llama 4 Scout 17B-16E Instruct109B, 17B active · 60.9 GB file | ~39–70 tok/sestimate | ~23–33 tok/sestimate |
| gpt-oss-120b117B, 5.1B active · 58.5 GB file | 111 tok/smedian of 1 run | ~85–124 tok/sestimate |
| Qwen3 Coder Next80B, 3B active · 45.2 GB file | 102 tok/smedian of 2 runs | ~128–187 tok/sestimate |
| Llama 3.3 70B Instruct70B · 39.6 GB file | ~21–26 tok/sestimate | ~11–17 tok/sestimate |
| Nemotron 3.5 Lightning 30B-A3B30B, 3B active · 23.7 GB file | ~157–279 tok/sestimate | ~91–133 tok/sestimate |
| Qwen3.6 35B-A3B35B, 3B active · 20.8 GB file | ~209–373 tok/sestimate | ~122–178 tok/sestimate |
| Ornith 1.5 35B-A3B35B, 3B active · 20.2 GB file | ~214–382 tok/sestimate | ~125–183 tok/sestimate |
| Qwen3 Coder 30B-A3B Instruct30.5B, 3.3B active · 17.3 GB file | ~199–354 tok/sestimate | ~116–169 tok/sestimate |
| Gemma 4 31B IT30.7B · 17.1 GB file | ~48–60 tok/sestimate | ~25–39 tok/sestimate |
| Granite 4.2 30B29.3B · 16.8 GB file | ~49–61 tok/sestimate | ~25–40 tok/sestimate |
| Qwen3.8 27B27B · 16.2 GB file | 50 tok/smedian of 4 runs | ~26–41 tok/sestimate |
| Gemma 4 26B-A4B IT25.2B, 3.8B active · 15.9 GB file | ~155–277 tok/sestimate | ~91–132 tok/sestimate |
| Qwen3.6 27B27B · 15.7 GB file | ~52–65 tok/sestimate | ~27–43 tok/sestimate |
| Devstral Small 2 24B (2512)24B · 13.3 GB file | ~61–77 tok/sestimate | ~32–50 tok/sestimate |
| gpt-oss-20b21B, 3.6B active · 10.8 GB file | ~200–357 tok/sestimate | ~117–171 tok/sestimate |
| Phi-414.7B · 8.3 GB file | ~99–123 tok/sestimate | ~52–81 tok/sestimate |
| Ministral 3 14B Reasoning (2512)13.9B · 7.7 GB file | ~107–133 tok/sestimate | ~56–87 tok/sestimate |
| Gemma 4 12B IT11.95B · 6.6 GB file | ~124–154 tok/sestimate | ~64–101 tok/sestimate |
| Qwen3.5 9B9B · 5.3 GB file | ~155–193 tok/sestimate | ~81–127 tok/sestimate |
| Granite 4.2 8B8.8B · 5.2 GB file | ~159–198 tok/sestimate | ~83–130 tok/sestimate |
| Gemma 4 E4B IT8B · 5.0 GB file | ~290–361 tok/sestimate | ~151–237 tok/sestimate |
| Qwen3 8B8.2B · 4.7 GB file | ~175–218 tok/sestimate | ~91–143 tok/sestimate |
| Llama 3.1 8B Instruct8B · 4.6 GB file | ~179–223 tok/sestimate | ~93–146 tok/sestimate |
| Qwen3.5 4B4B · 2.6 GB file | ~321–400 tok/sestimate | ~167–263 tok/sestimate |
Full pages: Apple Mac Studio (M5 Ultra, 80-core GPU) · 4x RTX 3090 (used) on EPYC 7003, open frame · check a different quant or context length