Running Uncensored DeepSeek-V4.1-Flash on 12GB VRAM
Researchers have released an uncensored quantization of DeepSeek-V4.1-Flash-1017 tailored for consumer hardware with 12GB VRAM using the Strata engine. By modifying 82 specific dense tensors per file while keeping routed experts intact, they reduced refusal rates from around 85% down to near zero without hurting benchmark performance. This breakthrough allows users with modest consumer hardware to run massive, highly capable language models locally while removing restrictive safety alignments. It demonstrates a surgical approach to model editing that preserves core reasoning capabilities while altering behavioral guardrails. The release builds upon ISTA-DASLab's GSQ-RCO quants and relies on an Engram table stored on an SSD alongside 64GB of system RAM. Testing showed that the IQ2_XXS variant achieved zero refusals while retaining performance on rigorous benchmarks like MMLU, GPQA-Diamond, and AIME25.
## BACKGROUND
DeepSeek-V4.1-Flash is an advanced Mixture-of-Experts (MoE) large language model that utilizes conditional memory architectures like Engram tables to offload massive parameter requirements to SSD storage. GSQ (Gumbel Softmax Quantization) and RCO (Riemannian Constrained Optimization) are model compression techniques developed by ISTA-DASLab that optimize individual bit-widths to preserve model accuracy.