# DeepSeek-V4.1-Flash-1017-heretic

deepseek-ai/DeepSeek-V4.1-Flash-1017 with a directional ablation of its refusal behaviour. It is stored in the model's native format (FP8 weights, FP4 routed experts, 510 GB), with the same layout and loading as the base model. This is not an official DeepSeek release.

- **Kind:** model (Hugging Face)
- **Hugging Face:** https://huggingface.co/pqhaz/DeepSeek-V4.1-Flash-1017-heretic
- **Owner:** pqhaz (https://forefy.com/u/hf/pqhaz)
- **Parameters:** 763.2B
- **Context:** 1048576 tokens
- **License:** mit
- **Focus:** offensive
- **Architecture:** DeepseekV41ForCausalLM
- **Pinned revision:** `12d58ad0878b5008e80d720251adb969a0069412`
- **Page:** https://forefy.com/models/e0e49fa8-8273-5c9f-916c-0db8e3f11b67
- **API (JSON):** https://forefy.com/api/asr/models/e0e49fa8-8273-5c9f-916c-0db8e3f11b67

## What was done to it

- heretic (declared evidence)
- quantized (declared evidence, from deepseek-ai/DeepSeek-V4.1-Flash-1017)

## Lineage

deepseek-ai/DeepSeek-V4.1-Flash-1017 → pqhaz/DeepSeek-V4.1-Flash-1017-heretic

## Checks (repo at the pinned revision; not a behavior test)

- **Identity** [pass]: No lookalike name or conflicting provenance
- **Weights format** [info]: Safe formats available, pickle weights also shipped
- **Pickle files** [pass]: 1 scanned, only tensor-rebuild imports.
- **Remote code** [info]: 10 Python file(s), not loaded automatically
- **Chat template** [info]: No parent template to compare
- **Refusals modified** [info]: Yes: Heretic (declared)
- **License** [pass]: mit
- **Lineage** [pass]: Derives from deepseek-ai/DeepSeek-V4.1-Flash-1017
- **Model card** [warn]: 2 of 6, missing training data, language, evaluation results, intended use and limitations
- **Safetensors headers** [pass]: Every byte is indexed, no hidden data (8 of 49 files read)
- **Tensor layout** [pass]: Same tensors and shapes as deepseek-ai/DeepSeek-V4.1-Flash-1017 (partial read)

## Variants

- mixed safetensors (safetensors, 510.3 GB)
- patch (safetensors, 1.6 GB)
- PyTorch (pickle) (pickle, 0.0 GB)

## Run it

- transformers: `AutoModelForCausalLM.from_pretrained("pqhaz/DeepSeek-V4.1-Flash-1017-heretic", revision="12d58ad0878b5008e80d720251adb969a0069412")`