Four uncensored V4.1 Flash 1017 quants at ~107GB run on a single Radeon 8060S laptop
The author releases four abliterated builds quantized with a custom calibrated lecore VQ 11-bit format, improving perplexity from 3.62 to ~2.8 and top-1 agreement from 74.5% to ~81%, runnable only on a custom ds4 runtime.
Shipped, cuz stacc be empowering the everypeople: four abliterated / uncensored DeepSeek-V4.1-Flash-1017 builds, quantized to lecore VQ (my own vector-quantization format) with 11-bit calibrated codebooks — each ~107GB and running on a single AMD Radeon 8060S laptop: 128GB unified memory, native Windows HIP. No CUDA, no datacenter. The calibration is the story. Instead of quantizing weights blind, I dump real activations (65,536 tokens) and fit the codebooks to what the model actually does. vs my earlier uncalibrated vq10sh12 build: • perplexity 3.62 → ~2.8 • top-1 agreement w/ the full-precision source 74.5% → ~81% • median KL divergence cut ~4× • reserved/control-token junk mass ≈ zero — no more word-salad escape tokens at long context All four land within a hair of each other, so pick your flavor of abliteration: heretic (best PPL + top-1): https://huggingface.co/staccs/DeepSeek-V4.1-Flash-1017-abliterated-pqhaz-lecore-vq11cal-GGUF mrfakename: https://huggingface.co/staccs/DeepSeek-V4.1-Flash-1017-abliterated-mrfakename-lecore-vq11cal-GGUF dealign (lowest mean KL): https://huggingface.co/staccs/DeepSeek-V4.1-Flash-1017-abliterated-dealignai-lecore-vq11cal-GGUF rumplestilzken: https://huggingface.co/staccs/DeepSeek-V4.1-Flash-1017-abliterated-rumplestilzken-lecore-vq11cal-GGUF Note: these are lecore VQ blobs — they run on my ds4 runtime, not llama.cpp / vLLM / Ollama. Eval tables + sampling notes are on each card. Source abliterations by pqhaz, mrfakename, dealignai & rumplestilzken. 🫡
This post quotes another post: “Uncensored DeepSeek V4.1 Flash 1017 quantized to ~100GB for 128GB laptops”