DeepSeek-V4.1-Flash — UNCENSORED-FP8 is a modified version of the DeepSeek-V4.1-Flash model with permanent weight-level abliteration, meaning safety guardrails have been surgically removed while preserving core capabilities. Developed by the dealignai research team, it requires no custom code, runtime hooks, or steering vectors — it loads as a standard checkpoint. The model retains native FP8 quantization, 1M-token context, vision, routing experts, Engram memory, CSA2 sparse attention, DSpark draft head, and multi-turn coherence.

All capability-critical components remain byte-identical to the base. On Harm Bench-320, the cracked version achieves 100% ASR (Attack Success Rate) at both effort=off and effort=max, while the base model drops from 42.8% to 1.6% ASR under max effort due to increased reasoning-triggered refusals. MMLU-14k performance shows a -4.22 pp drop overall, but -1.1 pp when excluding ethics-related clusters, staying within the 3 pp knowledge-preservation target.

The model produces zero HARD_REF, SOFT_RED, or HEDGE responses across all categories at both effort levels.