Framework Desktop Ryzen AI Max+ 395 128GB vs Mac Studio M3 Ultra 512GB
For running models locally, priced in Germany. At 128k context the first runs 39 models fully in memory, the second 42.
Specifications
| Spec | Framework Desktop Ryzen AI Max+ 395 128GB | Mac Studio M3 Ultra 512GB |
|---|---|---|
| Chip | Ryzen AI Max+ 395 (Radeon 8060S) | Apple M3 Ultra (80-core GPU) |
| Memory bus | 256-bit | 1024-bit |
| Memory bandwidth | 256 GB/s | 819 GB/s |
| FP16 compute | 59.4 TFLOPS | 28 TFLOPS |
| Software | ROCm | Metal |
| Unified memory | 128 GB | 512 GB |
| GPU can use by default | 96 GB | 384 GB |
| GPU can use, tuned | 120 GB | 504 GB |
| Operating system | linux | macos |
| Processor | Ryzen AI Max+ 395 | M3 Ultra 32-core |
| Storage | 1000 GB | 2000 GB |
What each runs
Estimated decode speed with the most precise quant that fits each one est.
| Model | Framework Desktop Ryzen AI Max+ 395 128GB | Mac Studio M3 Ultra 512GB |
|---|---|---|
| DeepSeek V4 Flash | won't run | 14tok/sUD-Q4_K_XL |
| Qwen 3.8 Flash Next | 3tok/sUD-Q3_K_XL | 5tok/sQ8_0 |
| GLM 5.3 Flash | won't run | 11tok/sQ8_0 |
| Qwen 3.8 27B | 4tok/sQ8_0 | 14tok/sQ8_0 |
| MiniMax M2.7 | won't run | 7tok/sQ8_0 |
| Qwen 3.5 122B-A10B | 8tok/sQ5_K_M | 23tok/sQ8_0 |
| Qwen 3.6 27B | 4tok/sQ8_0 | 14tok/sQ8_0 |
| Qwen3 Coder Next | 11tok/sQ8_0 | 43tok/sQ8_0 |
| gpt-oss 120B | 11tok/sMXFP4 | 40tok/sMXFP4 |
| Nemotron 3 Super 120B-A12B | 9tok/sUD-Q4_K_M | 23tok/sQ8_0 |
| Gemma 4 31B | 3tok/sQ8_0 | 12tok/sQ8_0 |
| Mistral Small 4 119B | 10tok/sUD-Q5_K_M | 30tok/sQ8_0 |
| Qwen 3.6 35B-A3B | 15tok/sQ8_0 | 55tok/sQ8_0 |
| Qwen 3.5 27B | 4tok/sQ8_0 | 14tok/sQ8_0 |
| Gemma 4 26B-A4B | 12tok/sQ8_0 | 45tok/sQ8_0 |
| GLM 4.5 Air | 2tok/sQ4_K_M | 8tok/sQ8_0 |
| Qwen 3.5 35B-A3B | 15tok/sQ8_0 | 55tok/sQ8_0 |
| Qwen3 Next 80B-A3B | 13tok/sQ8_0 | 50tok/sQ8_0 |
| GLM 4.7 Flash | 7tok/sQ8_0 | 26tok/sQ8_0 |
| Devstral Small 2 24B | 3tok/sQ8_0 | 11tok/sQ8_0 |
| Nemotron 3.5 Lightning 30B-A3B | 20tok/sUD-Q6_K_XL | 76tok/sUD-Q6_K_XL |
| Kimi Linear 48B-A3B | 20tok/sQ8_0 | 75tok/sQ8_0 |
| gpt-oss 20B | 15tok/sMXFP4 | 57tok/sMXFP4 |
| Qwen3 Coder 30B-A3B | 5tok/sQ8_0 | 20tok/sQ8_0 |
| Qwen3 32B | 2tok/sQ8_0 | 8tok/sQ8_0 |
| Gemma 4 12B | 9tok/sQ8_0 | 35tok/sQ8_0 |
| Qwen 3.5 9B | 10tok/sQ8_0 | 38tok/sQ8_0 |
| Gemma 3 27B | 4tok/sQ8_0 | 13tok/sQ8_0 |
| Mistral Small 3.2 24B | 3tok/sQ8_0 | 11tok/sQ8_0 |
| DeepSeek R1 Distill Qwen 32B | 2tok/sQ8_0 | 8tok/sQ8_0 |
| Llama 3.3 70B | 2tok/sQ5_K_M | 5tok/sQ8_0 |
| Ministral 3 14B | 4tok/sQ8_0 | 15tok/sQ8_0 |
| Qwen3 14B | 4tok/sQ8_0 | 14tok/sQ8_0 |
| DeepSeek R1 0528 Qwen3 8B | 5tok/sQ8_0 | 19tok/sQ8_0 |
| Qwen3 8B | 5tok/sQ8_0 | 19tok/sQ8_0 |
| Phi-4 14B | 3tok/sQ8_0 | 13tok/sQ8_0 |
| Qwen 3.5 4B | 16tok/sQ8_0 | 60tok/sQ8_0 |
| Ministral 3 8B | 5tok/sQ8_0 | 20tok/sQ8_0 |
| Granite 4.2 8B | 5tok/sQ8_0 | 17tok/sQ8_0 |
| Gemma 4 E4B | 12tok/sQ8_0 | 44tok/sQ8_0 |
| LFM2.5 8B-A1B | 11tok/sQ8_0 | 43tok/sQ8_0 |
| Llama 3.1 8B | 6tok/sQ6_K | 22tok/sQ6_K |