Abliterated DeepSeek V4.1 Flash 1017: capability intact, 256GB Mac MLX build
Five observations: most users may not need it, benchmark scores stay within noise of the original, differences show on edge cases, an MLX build runs on a 256GB Mac, and some endpoints charge 3–4x.
i’ve been looking at the abliterated DeepSeek V4.1 Flash 1017 release. 5 things stood out: 1. i’m not sure most people even need it. regular DeepSeek already refuses far less than Claude on a lot of technical work. 2. the interesting part is that abliteration doesn’t seem to hurt capability much. benchmark scores are reportedly basically identical to the original within sampling noise. 3. the real difference shows up on edge cases. tasks that normal DeepSeek still refuses can go through on the abliterated version without the same friction. 4. local is possible, but not cheap. there’s already an MLX build that can run on a 256GB Mac, which puts this in serious workstation territory. 5. provider pricing can kill the value fast. some abliterated endpoints are charging around 3–4x the normal rate. this is the same DeepSeek with fewer brakes.