Mac Studio M3 Ultra 96GB vs Framework Desktop Ryzen AI Max+ 395 128GB
For running models locally, priced in Germany. At 32k context the first runs 38 models fully in memory, the second 41.
Specifications
| Spec | Mac Studio M3 Ultra 96GB | Framework Desktop Ryzen AI Max+ 395 128GB |
|---|---|---|
| Chip | Apple M3 Ultra (80-core GPU) | Ryzen AI Max+ 395 (Radeon 8060S) |
| Memory bus | 1024-bit | 256-bit |
| Memory bandwidth | 819 GB/s | 256 GB/s |
| FP16 compute | 28 TFLOPS | 59.4 TFLOPS |
| Software | Metal | ROCm |
| Unified memory | 96 GB | 128 GB |
| GPU can use by default | 72 GB | 96 GB |
| GPU can use, tuned | 88 GB | 120 GB |
| Operating system | macos | linux |
| Processor | M3 Ultra 32-core | Ryzen AI Max+ 395 |
| Storage | 1000 GB | 1000 GB |
What each runs
Estimated decode speed with the most precise quant that fits each one est.
| Model | Mac Studio M3 Ultra 96GB | Framework Desktop Ryzen AI Max+ 395 128GB |
|---|---|---|
| DeepSeek V4 Flash | won't run | 9tok/sUD-Q2_K_XL |
| Qwen 3.8 Flash Next | won't run | 3tok/sUD-Q3_K_XL |
| Qwen 3.8 27B | 17tok/sQ8_0 | 5tok/sQ8_0 |
| MiniMax M2.7 | won't run | 7tok/sUD-Q2_K_XL |
| Qwen 3.5 122B-A10B | 58tok/sQ3_K_M | 9tok/sQ6_K |
| Qwen 3.6 27B | 17tok/sQ8_0 | 5tok/sQ8_0 |
| Qwen3 Coder Next | 78tok/sQ6_K | 17tok/sQ8_0 |
| gpt-oss 120B | 74tok/sMXFP4 | 20tok/sMXFP4 |
| Nemotron 3 Super 120B-A12B | 50tok/sUD-Q3_K_M | 10tok/sUD-Q4_K_M |
| Gemma 4 31B | 15tok/sQ8_0 | 4tok/sQ8_0 |
| Mistral Small 4 119B | 62tok/sUD-Q4_K_M | 13tok/sUD-Q6_K |
| Qwen 3.6 35B-A3B | 84tok/sQ8_0 | 22tok/sQ8_0 |
| Qwen 3.5 27B | 17tok/sQ8_0 | 5tok/sQ8_0 |
| Gemma 4 26B-A4B | 63tok/sQ8_0 | 17tok/sQ8_0 |
| GLM 4.5 Air | 21tok/sQ3_K_M | 4tok/sQ5_K_M |
| Qwen3 Next 80B-A3B | 97tok/sQ6_K | 21tok/sQ8_0 |
| Qwen 3.5 35B-A3B | 84tok/sQ8_0 | 22tok/sQ8_0 |
| GLM 4.7 Flash | 45tok/sQ8_0 | 12tok/sQ8_0 |
| Devstral Small 2 24B | 18tok/sQ8_0 | 5tok/sQ8_0 |
| Nemotron 3.5 Lightning 30B-A3B | 89tok/sUD-Q6_K_XL | 24tok/sUD-Q6_K_XL |
| gpt-oss 20B | 99tok/sMXFP4 | 26tok/sMXFP4 |
| Kimi Linear 48B-A3B | 91tok/sQ8_0 | 24tok/sQ8_0 |
| Qwen3 Coder 30B-A3B | 50tok/sQ8_0 | 13tok/sQ8_0 |
| Qwen3 32B | 12tok/sQ8_0 | 3tok/sQ8_0 |
| Gemma 4 12B | 39tok/sQ8_0 | 10tok/sQ8_0 |
| Qwen 3.5 9B | 50tok/sQ8_0 | 13tok/sQ8_0 |
| Mistral Small 3.2 24B | 18tok/sQ8_0 | 5tok/sQ8_0 |
| Gemma 3 27B | 17tok/sQ8_0 | 4tok/sQ8_0 |
| DeepSeek R1 Distill Qwen 32B | 12tok/sQ8_0 | 3tok/sQ8_0 |
| Llama 3.3 70B | 9tok/sQ5_K_M | 2tok/sQ8_0 |
| Ministral 3 14B | 27tok/sQ8_0 | 7tok/sQ8_0 |
| Qwen3 14B | 25tok/sQ8_0 | 7tok/sQ8_0 |
| DeepSeek R1 0528 Qwen3 8B | 39tok/sQ8_0 | 10tok/sQ8_0 |
| Qwen3 8B | 39tok/sQ8_0 | 10tok/sQ8_0 |
| Phi-4 14B | 24tok/sQ8_0 | 6tok/sQ8_0 |
| Qwen 3.5 4B | 95tok/sQ8_0 | 25tok/sQ8_0 |
| Ministral 3 8B | 39tok/sQ8_0 | 10tok/sQ8_0 |
| Granite 4.2 8B | 36tok/sQ8_0 | 10tok/sQ8_0 |
| Gemma 4 E4B | 58tok/sQ8_0 | 15tok/sQ8_0 |
| LFM2.5 8B-A1B | 119tok/sQ8_0 | 32tok/sQ8_0 |
| Llama 3.1 8B | 49tok/sQ6_K | 13tok/sQ6_K |
| GLM 5.3 Flash | won't run | won't run |