Delta Inference Engine v0.93 expands local DeepSeek inference support
The developer released Delta Inference Engine v0.93 with support for 64, 96 and 128 GB of RAM, Unsloth quantization formats, EXL3 and basic multi-GPU configurations. On one Windows system with an RTX 3090, the reported Unsloth UD-IQ2_XXS configuration reached 41 tokens per second with 192 GB RAM and scored 57 out of 99 on the developer’s benchmark, compared with 52 for the previous default quantization and 59 for the full model via API. The update also includes public CI builds, faster prompt handling and fixes for long contexts and large model files.
The release broadens the hardware and quantization options for running DeepSeek-V4.1-Flash locally, but its performance and quality figures come from one developer test system and still need validation on other hardware.
#Delta Inference Engine v0.93#DeepSeek-V4.1-Flash-1017#Unsloth#EXL3#Strata#Xolotl#Project Maya#GitHub Actions