DGX Spark vs Framework Desktop 128 GB for local AI
Today the Framework Desktop 128 GB is cheaper: $3,464.65 against $7,799.98 for the DGX Spark. Per gigabyte of memory the Framework Desktop 128 GB wins, $27 against $60. For Qwen3.8 27B · Q4_K_M the DGX Spark is faster, about 11 against 7.8 tokens per second.
| NVIDIA DGX Spark | Framework Desktop, Ryzen AI Max+ 395, 128 GB | |
|---|---|---|
| Lowest price today, delivered | $7,799.98 on eBay | $3,464.65 on eBay |
| Price per GB | $60/GB | $27/GB |
| Memory | 128 GB LPDDR5x unified | 128 GB LPDDR5x-8000 unified |
| Memory bandwidth | 273 GB/s | 256 GB/s |
| FP16 compute, dense | 125 TFLOPS | 29.7 TFLOPS |
| Reference models that fit at Q4_K_M | 18 of 28 | 18 of 28 |
| Power | 240 W | 140 W |
| Form factor | Desktop box, 150 x 150 x 50 mm | Mini ITX desktop, 4.5 L |
| Interface | unified memory | unified memory |
| Software platform | cuda | rocm |
| Launched | 2025 | 2025 |
| Launch price | $3,999 | $1,999 |
Qwen3.8 27B · Q4_K_M by quantization
| Quant | DGX Spark | Framework Desktop 128 GB |
|---|---|---|
| FP16 | ~3.3 tok/s | ~2.4 tok/s |
| Q8_0 | ~6.2 tok/s | ~4.5 tok/s |
| Q6_K | ~8.0 tok/s | ~5.8 tok/s |
| Q5_K_M | ~9.3 tok/s | ~6.7 tok/s |
| Q4_K_M | ~11 tok/s | ~7.8 tok/s |
| Q3_K_M | ~13 tok/s | ~9.7 tok/s |
| IQ2_M | ~19 tok/s | ~14 tok/s |
Fit means the weights plus an 8K context cache fit in memory; speed is the estimated decode rate. Prices are the cheapest live eBay listing that passed our filters, item plus shipping. How we estimate.