Summary
User reports DeepSeek V4.1 Flash at 11-18 tok/s on an RTX 4090 with 96GB of DDR5-6800 system RAM and a 4TB PCIe 5 NVMe drive.
Setup is Strata with an IQ2_XXS quant; the model runs partly from system memory.
User calls the speed quite good for the model but says prefill is very bad.