DeepSeek v4.1 Flash Uncensored

DeepSeek-V4.1-Flash UNCENSORED-FP8 removes all safety guardrails while retaining the base model’s MMLU performance, vision, reasoning, 1 M-token context, and multi-turn coherence. The checkpoint loads like the original model and achieves 100% refusal-free responses on HarmBench-320, compared with the guarded version’s 42.8% refusal rate at maximum effort. It can be deployed via Transformers, vLLM, SGLang, Docker, or Docker Model Runner using standard commands.