Uncensored DeepSeek-V4.1 quants cut refusals from 85% to 0% on 12 GB VRAM
gladkos · reddit · 2026-11-02
The AtomicChat team released uncensored GGUF quants of DeepSeek-V4.1-Flash-1017, built on ISTA-DASLab's GSQ-RCO quants and runnable on a 12 GB GPU + 64 GB RAM + SSD via the Strata Engine.
The edit is surgical: only 82 dense tensors per file are touched (attention output, shared expert down proj, Engram value proj), while the 384 routed experts remain byte-for-byte identical to ISTA's. Headers and file sizes match, so users can swap shards and keep the original Engram file.
JailbreakBench results (stock → modified, same Strata setup): IQ1M refusals 88.9%→0% (in-thinking 27/30→1/30), IQ2XXS 85.2%→0% (in-thinking 27→0), IQ2XS 86.4%→0% (still 2/30 in thinking). Capability checks on IQ2XXS: GPQA-D 83.3% (unchanged), AIME25 80.0% vs 76.7%, with KL vs stock around 0.02.
The authors note they only tested 8K context on a rented GPU and haven't verified long context or real-world speed on a 12 GB card. A llama.cpp-compatible earlier release is also available. Model and full metrics are on Hugging Face.
More from Infra
- Mistral Large 4 tech report: 1T-param MoE, 606GB quantized, runs on four GB300 GPUs — _AndrewZhao · 2026-11-03
- Qualcomm hits a significant milestone on its path into the data center — karlfreund · 2026-11-03
- Anthropic's Q3 growth slows to 47% QoQ from 144%, against $895.4B compute commitments — petrusenko_max · 2026-11-03
- vLLM adds Rubin dual-die localization: pairing MoE weight shards with local compute on both dies — anbarth · 2026-11-03
- NVIDIA reportedly producing ~1,000 GB300 racks per week; 1.2M GPUs bound for one campus — acmoytoy · 2026-11-03
- "The VC welfare program is over": Redditor defends OpenAI usage cuts with compute math — R3K4CE · 2026-11-03