It freaking worked lolπŸ”₯ Deepseek-v4-flash-0731 @UnslothAI 's IXQ2/Q3 checkpoint on one single RTX4090 with just 64 GB of RAM at usable token rate without dspark. All kernels running on Blaze (my custom developed ML compiler + inference engine) - no llama.cpp or vllm in the picture. The setup keeps heavily utilized experts in RAM with CPU (a trick from this guy ) Current tps is around 8 tps with slight expert miss causing a disk read which lowers it to 5 tps momentarily. With dspark and p...