A configuration of two DGX Spark units and a Connect-X7 cable provides approximately 250GB of usable memory for $7,000 USD, enabling the use of large models at 4-bit quantization. This setup supports a recent wave of "miniboss" models that fit within this memory constraint.
- Tencent Hy3 (182GB)
- StepFun 3.7 Flash (129GB)
- MiniMax M2.7 (131GB)
- Deepseek V4 Flash (160GB)
- Xiaomi MiMo 2.5 (181GB)
The author suggests testing these models on OpenRouter for $7 before committing to the hardware investment.