Framework Desktop Ryzen AI Max+ 395 128GB vs NVIDIA DGX Spark 128GB
For running models locally, priced in Germany. At 8k context the first runs 41 models fully in memory, the second 42.
Specifications
| Spec | Framework Desktop Ryzen AI Max+ 395 128GB | NVIDIA DGX Spark 128GB |
|---|---|---|
| Chip | Ryzen AI Max+ 395 (Radeon 8060S) | GB10 Grace Blackwell |
| Memory bus | 256-bit | 256-bit |
| Memory bandwidth | 256 GB/s | 273 GB/s |
| FP16 compute | 59.4 TFLOPS | 62 TFLOPS |
| Software | ROCm | CUDA |
| Unified memory | 128 GB | 128 GB |
| GPU can use by default | 96 GB | 119 GB |
| GPU can use, tuned | 120 GB | 122 GB |
| Operating system | linux | dgx-os |
| Processor | Ryzen AI Max+ 395 | 20-core Arm (GB10) |
| Storage | 1000 GB | 4000 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 | NVIDIA DGX Spark 128GB |
|---|---|---|
| DeepSeek V4 Flash | 11tok/sUD-Q2_K_XL | 14tok/sUD-Q2_K_XL |
| Qwen 3.8 Flash Next | 3tok/sUD-Q3_K_XL | 3tok/sUD-Q4_K_XL |
| GLM 5.3 Flash | won't run | 12tok/sUD-Q2_K_XL |
| MiniMax M2.7 | 15tok/sUD-Q2_K_XL | 15tok/sUD-Q3_K_M |
| Qwen 3.8 27B | 5tok/sQ8_0 | 6tok/sQ8_0 |
| Qwen 3.5 122B-A10B | 10tok/sQ6_K | 12tok/sQ6_K |
| Qwen 3.6 27B | 5tok/sQ8_0 | 6tok/sQ8_0 |
| gpt-oss 120B | 25tok/sMXFP4 | 30tok/sMXFP4 |
| Qwen3 Coder Next | 19tok/sQ8_0 | 23tok/sQ8_0 |
| Nemotron 3 Super 120B-A12B | 10tok/sUD-Q4_K_M | 9tok/sUD-Q6_K |
| Gemma 4 31B | 4tok/sQ8_0 | 5tok/sQ8_0 |
| Mistral Small 4 119B | 14tok/sUD-Q6_K | 13tok/sQ8_0 |
| Qwen 3.6 35B-A3B | 26tok/sQ8_0 | 31tok/sQ8_0 |
| Qwen 3.5 27B | 5tok/sQ8_0 | 6tok/sQ8_0 |
| Gemma 4 26B-A4B | 19tok/sQ8_0 | 22tok/sQ8_0 |
| GLM 4.5 Air | 5tok/sQ6_K | 5tok/sQ8_0 |
| Qwen 3.5 35B-A3B | 26tok/sQ8_0 | 31tok/sQ8_0 |
| Qwen3 Next 80B-A3B | 25tok/sQ8_0 | 30tok/sQ8_0 |
| GLM 4.7 Flash | 15tok/sQ8_0 | 18tok/sQ8_0 |
| Devstral Small 2 24B | 5tok/sQ8_0 | 6tok/sQ8_0 |
| Nemotron 3.5 Lightning 30B-A3B | 25tok/sUD-Q6_K_XL | 30tok/sUD-Q6_K_XL |
| gpt-oss 20B | 32tok/sMXFP4 | 39tok/sMXFP4 |
| Kimi Linear 48B-A3B | 26tok/sQ8_0 | 31tok/sQ8_0 |
| Qwen3 Coder 30B-A3B | 21tok/sQ8_0 | 25tok/sQ8_0 |
| Qwen3 32B | 4tok/sQ8_0 | 5tok/sQ8_0 |
| Gemma 4 12B | 11tok/sQ8_0 | 13tok/sQ8_0 |
| Qwen 3.5 9B | 14tok/sQ8_0 | 17tok/sQ8_0 |
| Mistral Small 3.2 24B | 5tok/sQ8_0 | 6tok/sQ8_0 |
| Gemma 3 27B | 5tok/sQ8_0 | 6tok/sQ8_0 |
| DeepSeek R1 Distill Qwen 32B | 4tok/sQ8_0 | 5tok/sQ8_0 |
| Llama 3.3 70B | 2tok/sQ8_0 | 2tok/sQ8_0 |
| Ministral 3 14B | 9tok/sQ8_0 | 11tok/sQ8_0 |
| Qwen3 14B | 8tok/sQ8_0 | 10tok/sQ8_0 |
| DeepSeek R1 0528 Qwen3 8B | 14tok/sQ8_0 | 17tok/sQ8_0 |
| Qwen3 8B | 14tok/sQ8_0 | 17tok/sQ8_0 |
| Phi-4 14B | 8tok/sQ8_0 | 10tok/sQ8_0 |
| Qwen 3.5 4B | 29tok/sQ8_0 | 35tok/sQ8_0 |
| Ministral 3 8B | 14tok/sQ8_0 | 17tok/sQ8_0 |
| Granite 4.2 8B | 13tok/sQ8_0 | 16tok/sQ8_0 |
| Gemma 4 E4B | 17tok/sQ8_0 | 20tok/sQ8_0 |
| LFM2.5 8B-A1B | 58tok/sQ8_0 | 69tok/sQ8_0 |
| Llama 3.1 8B | 18tok/sQ6_K | 22tok/sQ6_K |