Community Enables Running Uncensored DeepSeek on 12GB VRAM Devices

On November 2, 2026 at 23:52, the Reddit community LocalLLaMA posted about running an "uncensored" DeepSeek-V4.1-Flash on a 12GB GPU. The approach builds on the DeepSeek-V4.1-Flash-1017 GSQ-RCO quants produced by ISTA-DASLab, the same ones typically run in Strata Engine with a 12 GB GPU and the Engram table stored on SSD. According to the post, the refusals were removed by touching only 82 dense tensors per file: the attention output, the shared expert down proj, and the Engram value proj. The 384 routed experts are claimed to be byte for byte identical to ISTA's, so their GSQ rounding is kept, and headers and file sizes are unchanged. The reported measurement is a drop in refusal rate from 85% to 0%, even in thinking mode. The setup is described as local quantized inference via Strata Engine on a 12 GB GPU VRAM with 64GB RAM and an SSD, across 3 sizes. At this point the event rests on this single community post, with no additional reports or official response.