DeepSeek-V4.1-Flash-Abliterated

DeepSeek-V4.1-Flash-Abliterated

<div align="center"> <img src="https://github.com/deepseek-ai/DeepSeek-V2/blob/main/figures/logo.svg?raw=true" width="50%" alt="DeepSeek" /> <p><a href="https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash">Original Model</a> · <a href="https://github.com/p-e-w/heretic">Heretic</a> · <a href="LICENSE">MIT License</a></p> </div>

Introduction

This is a community abliterated conversion of DeepSeek-V4.1-Flash, published by s-zaizen. It is not an official DeepSeek release.

The checkpoint retains the source model's architecture, tokenizer, vision components, Engram memory, MTP/DSpark tensors, native context configuration, and original FP8/MXFP4 storage formats. The release contains 48 safetensors shards totaling 510.297 GB.

Abliteration

Created with p-e-w/heretic at commit 3521f8648a0dccf6e12a92666862632235fac7e6. A native DeepSeek V4.1 compatibility adapter uses Heretic's residual-direction construction and heretic.model.Model.abliterate API. The selected per_all_4p5 intervention applies projected per-layer directions to all 40 attention output projections with full row normalization and rank-3 norm-preserving LoRA merging. It does not modify expert, Engram, vision, embedding, or MTP/DSpark tensors.

Refusals

Both values use the same actual Heretic CLI evaluation over 100 examples from its configured mlabonne/harmful_behaviors test split, with max_response_length=100, batch_size=4, temperature 0, chat thinking mode, and seed 42. The bold value is the post-conversion result. Prompt and response text was not logged or published. This refusal keyword rate is not a general capability benchmark.

Inference

Download the checkpoint:

hf download s-zaizen/DeepSeek-V4.1-Flash-Abliterated --local-dir DeepSeek-V4.1-Flash-Abliterated

Use a runtime that supports the original DeepSeek V4.1 FP8/MXFP4 checkpoint format and prompt encoding. The upstream reference inference and encoding files are included unchanged.

License

The original DeepSeek MIT license is retained for the model and accompanying source files. Heretic is credited as the abliteration tool and is distributed under its own AGPL-3.0 license.

Credits

DeepSeek — original model and inference code. Philipp Emanuel Weidmann and Heretic contributors — abliteration tooling. s-zaizen — conversion and publication.