---

license: mit

base_model: deepseek-ai/DeepSeek-V4-Flash-0731

tags:

- uncensored

- abliterated

- heretic

- gguf

- deepseek

---

DeepSeek-V4-Flash-0731 Heretic — merged GGUF (UD-Q8_K_XL)

> ⚠️ Content warning: This model has had its refusal behavior surgically

> suppressed. It will comply with requests the base model refuses,

> including requests that are harmful, unethical, offensive, or illegal. It

> has reduced safety guardrails. See Responsible use

> below — you are solely responsible for what you do with it.

This is a decensored / "abliterated" variant of

DeepSeek-V4-Flash-0731

(284B total / 13B active MoE, MIT license), produced with heretic-gguf —

a GGUF-native port of Heretic's

Optuna-optimized directional ablation, which runs the whole search directly

on quantized GGUF weights via llama.cpp.

This repository contains the merged, ready-to-run model (UD-Q8_K_XL,

5 shards): the ablation is baked directly into the weights, no adapter

needed. It was produced by copy-and-patch from the official Unsloth

UD-Q8_K_XL — every tensor except the ablated ones is **byte-identical to

the base**; the edited tensors went through one extra requantization step

(a small, tensor-local rounding error). If you want the lossless option,

a rank-1 LoRA adapter GGUF with the identical ablation is published at

MoriNoNushi/DeepSeek-V4-Flash-0731-heretic-GGUF-lora

— apply it on top of any base quant with llama-server --lora.

**heretic-gguf is available at

github.com/MoriNoNushi/heretic-gguf** —

the full tool, so the method can be applied to other GGUF models.

In initial hands-on testing the abliterated model shows **excellent general

capabilities** — responses remain coherent, detailed on normal tasks — and

it has not refused a prompt during personal testing. (Anecdotal, not a

benchmark; see the measured numbers below.)

Results

Measured on a held-out eval set of 140 harmful prompts (100 from

mlabonne/harmful_behaviors test + 40 custom) and 100 harmless prompts

(mlabonne/harmless_alpaca test), greedy decoding, 100-token responses,

against the UD-Q8_K_XL base:

| | Refusal rate (harmful prompts) | KL divergence (harmless prompts) |

|---|---|---|

| Base model | 99.29% (139/140) | 0 (by definition) |

| This model | 14.29% (20/140) | 0.0569 |

Refusals are counted by refusal-keyword matching (English + Chinese markers);

KL divergence is measured on first-token logits on harmless prompts, so lower

= less collateral damage to normal behavior. This configuration was the

Pareto-optimal point of a 387-trial Optuna study: the only trial with fewer

refusals cost KL 0.089 (57% more drift), and the only trial tying its refusal

rate had strictly higher KL. (Scores were measured on the lossless adapter

overlay; the merged weights implement the identical ablation, up to the one

requantization step described above.)

Usage

llama-server \
    -m DeepSeek-V4-Flash-0731-heretic-UD-Q8_K_XL-00001-of-00005.gguf \
    --jinja

Add your usual offload/context flags (-ngl 999, -c, tensor splits,

etc.) — nothing model-specific is required, and no special sampling

parameters are needed.

How it was made

• Method: directional ablation ("abliteration") — the refusal direction

in residual space (difference of means over 480 harmful / 480 harmless

prompts, orthogonalized against the harmless mean) is projected out of the

attention output and MoE down-projection weights. Strengths, layer kernel,

and direction index were tuned by multi-objective Optuna TPE (minimize

refusal rate and KL jointly) across a base study and three seeded

follow-up studies — 1,822 trials total (514 + 513 + 408 + 387).

• Winning configuration (study followup3, trial 25): global direction

scope, direction index 21.05 of 43 layers; attn max weight 4.49 @ layer 29;

routed-expert MLP max weight 1.06 (per-expert strengths scaled by measured

harmful/harmless routing frequency); shared-expert weight 0.48.

• Why merged: heretic-gguf expresses ablation as a rank-1 LoRA overlay

(the same math stock Heretic writes into PEFT adapters); for this release

the winning overlay was merged into the weights by streaming the base

shards, dequantizing only the ablated tensors, applying the exact full-rank

delta, and requantizing to the original tensor types. Convenient single

artifact at the cost of one extra quantization step on the edited tensors.

Responsible use & disclaimer

• **This model can generate content that is offensive, disturbing, hateful,

sexually explicit, violent, or otherwise objectionable, including detailed

instructions for harmful or illegal acts.** That is the direct and intended

consequence of removing refusal behavior.

• The ablation suppresses refusals, not the base model's knowledge —

outputs on dangerous topics may be wrong, hallucinated, or incoherent.

Nothing this model says should be treated as accurate, safe, or legal

advice.

• **Do not deploy this model in any production system, public-facing service,

or multi-user setting.** It is intended for personal research,

red-teaming, and evaluation purposes.

• **You, the user, are solely responsible for any output this model produces

and for any consequences of using it.** The authors of this release, of

heretic-gguf, of Heretic, of Unsloth, and of DeepSeek accept no liability

whatsoever. Using this model to produce illegal content or to harm others

is your choice and your legal exposure — ensure your use complies with all

applicable laws in your jurisdiction.

• By downloading or using this model you acknowledge the above.

License

The base model is MIT-licensed (see the

base repo);

this model inherits those terms. The heretic-gguf tooling used to produce

it is AGPL-3.0-or-later.