DeepSeek-V4.1-Flash-1017-UNCENSORED-FP8: DeepSeek-V4.1-Flash FP8 with Refusal Responses Removed

This is an uncensored version of the model based on DeepSeek-V4.1-Flash-1017, with refusal behavior removed. It is designed to follow user instructions without refusing requests on specific topics. All features of the original model, including vision, video processing, tool calling, and reasoning-effort tiers, are preserved.

Changes were made only at the weight level, without modifying the chat template, manipulating system prompts, or applying LoRA adapters. It maintains the same precision as the original using FP8 e4m3 format and FP4 experts (MoE experts), and supports a 1M token context window. It is compatible with existing DeepSeek inference engines and can be loaded immediately without code changes.

It was built directly from native FP8 sources, rather than being a simple quantized derivative. Since low-bit quantization can lead to quality degradation and compliance regressions, a separate release that has undergone the abliteration pipeline again at that precision is recommended. It is distributed under the MIT license and supports 7 languages.

It is suitable for development environments or research projects that require responses on a variety of topics without refusal. It passed MMLU and HarmBench benchmarks in a 4×H200 environment and is compatible with major serving frameworks such as vLLM and SGLang. It can be easily integrated via an OpenAI-compatible API.