DeepSeek V4.1 Flash Uncensored: weight-level abliteration and test results

Dealignai published an FP8 checkpoint of DeepSeek V4.1 Flash with refusal behavior removed in the weights. On HarmBench-320 the model answers 100% of prompts versus 1.6% for the base model at max thinking, while MMLU drops from 87.0% to 82.7%.

• HarmBench-320: 100% answered vs 1.6% for the base model at thinking max
• MMLU fell from 87.0% to 82.7%, with moral scenarios dropping 76.9% to 37.0%
• 552B MoE with 1M input and 131,072 output context, hosted only
• Priced at about 22,500 credits per 1M input and 90,000 per 1M output tokens