Community
Delta Inference Engine v0.93 runs DeepSeek-V4.1-Flash-1017 on one RTX 3090 with new Unsloth quants
The update adds 64, 96 and 128 GB RAM support and reports 41 tok/s at 192 GB with the UD-IQ2_XXS quant, with first token about 40% sooner. The author's own 99-question benchmark scored 57 for this quant against 59 for the full API model.