DeepSeek V4.1 Flash 1017 Uncensored (HeimdaLLM) (consensusprotocol)

This entry points to a third-party uncensored build of the DeepSeek V4.1 Flash 1017 checkpoint, published on Hugging Face by the account @dealignai with compute contributed by @jordanschenck. Rather than a fresh training run, the release is produced by applying an abliteration edit to the base weights, stripping refusal behavior from the original model so it answers more directly and less defensively. The checkpoint is distributed in native FP8 e4m3 format with ue8m0 block scales and FP4 mixture-of-experts experts across a 40-layer MoE stack, and the publisher asks downstream users to credit the repository when redistributing, fine-tuning, or benchmarking it.

The publisher's most useful guidance concerns precision: quantizing this FP8 bundle straight down to low bit-depths stacks quantization error on top of the abliteration edit, and compliance can regress at low bits, so the recommended path is to re-run the abliteration pipeline natively at the target precision instead of converting the FP8 file after the fact. A sibling EXL3 build at roughly 2.9 bits per weight is offered separately for that reason, and the model card recommends a temperature of 1.0 with top-p of 1.0 and a large maximum token allowance for its heaviest reasoning mode. Because the abliteration removes safeguards, this build suits users who want an open-ended assistant for creative or sensitive tasks and can handle the precision and sampling choices carefully.