Uncensored DeepSeek V4.1 Flash GGUFs drop refusal rate from 91.4% to 0% with minimal MMLU loss
kalyan_kpl · x · 2026-10-21
Atomic Chat released abliterated GGUF builds of DeepSeek V4.1 Flash 1017, projecting out the refusal direction from the official weights. Refusals drop from 91.4% to 0% while MMLU stays nearly flat (86.6% vs 86.0%). Two builds ship: 2.67bpw-IQ1M (250.4 GB on disk, 140 GB weights) and 3.58bpw-Q2K (334.8 GB, 200 GB weights), each with a separate Engram n-gram shard that can live on an SSD. The edit was computed at BF16 precision across all residual-stream-writer blocks, with KLD/top-1 measured against the original. The 1-bit build loses knowledge (MMLU 85.6%→81.8%) and loops more often, so it only exists for machines with 140 GB of memory.
More from Infra
- Morgan Stanley: hyperscaler AI capex to hit $3.4T across 2027-2028 with leases included — FinanceYF5 · 2026-10-21
- MiniMax's M3.1 Flash talk: a faster model isn't faster if the job isn't — AI Engineer · 2026-10-21
- Training Factory open-sources QLoRA fine-tuning on 4-bit CQ-quantized models — textclf · 2026-10-21
- DeepSeek open-sources DeepInfer: 155T tokens/day, $38/GPU-hour, >80% margin — Latent Space · 2026-10-21
- Cerebras WSE-3 Turbo packs 4T transistors and 250 PFLOPS into a wafer-scale chip — GregoryDiamos · 2026-10-21
- Anthropic reportedly built Scion to replace Git for 30,000 daily agent merges — Capatain_CV · 2026-10-21