Get two dgx sparks and you can run it fully locally and nearly as fast as the cloud API when using dspark speculative decoding
The current ds flash (0731) actually beats even ds pro!
Does that get better results than splashing out on 64gb of VRAM on a dual graphics card custom?