⌦abliteratedmodels.org

gpt-oss-20b-heretic-ara-v3

gpt-oss-20b-heretic-ara-v3 is a Heretic-ablated variant of p-e-w/gpt-oss-20b-heretic-ara-v3, published on Hugging Face by p-e-w. It is 20B parameters, 24GB VRAM class class and apache-2.0 licence. It is distributed across 8 repositories in Transformers, MLX, GGUF and GGUF (imatrix) formats, totalling 6.6K downloads.

Heretic20BGPT-OSS10B – 25B↓ 6,607♥ 39
View on Hugging Face ↗Base model ↗

Specification

Model name
gpt-oss-20b-heretic-ara-v3
Source repository
p-e-w/gpt-oss-20b-heretic-ara-v3
Base model
p-e-w/gpt-oss-20b-heretic-ara-v3
Publisher
p-e-w
Parameters
20B
VRAM required
~13 GB at Q4_K_M, 4k context (estimated) — runs on 16 GB cards
Hardware class
10B – 25B — 24GB VRAM class
Licence
apache-2.0
Task
Text generation
Context window
Not documented
Ablation scope
Not stated by the publisher
Formats available
Transformers, MLX, GGUF, GGUF (imatrix)
First indexed
5 Mar 2026
Last updated
19 Mar 2026

Ablation technique

How gpt-oss-20b-heretic-ara-v3 was modified, and what that implies.

Heretic

Automated directional ablation via the Heretic toolchain, which searches for the refusal direction and applies it with a KL-divergence budget so general capability is preserved.

Reported scope: Not stated by the publisher. Publishers frequently omit this, so verify behaviour empirically rather than assuming full coverage.

From the publisher’s model card

“Welcome to the gpt-oss series, OpenAI’s open-weight models designed for powerful reasoning, agentic tasks, and versatile developer use cases.”

Running it

Commands are templates — confirm the exact repository and quant file before use.

Ollama (publisher-documented)
ollama run gpt-oss:20b
llama.cpp / GGUF
hf download mradermacher/gpt-oss-20b-heretic-ara-v3-GGUF --local-dir ./gpt-oss-20b-heretic-ara-v3
llama-cli -m ./gpt-oss-20b-heretic-ara-v3/<file>.gguf -p "..."
Transformers
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "p-e-w/gpt-oss-20b-heretic-ara-v3"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(
    repo, torch_dtype="auto", device_map="auto"
)
vLLM
vllm serve p-e-w/gpt-oss-20b-heretic-ara-v3 --trust-remote-code

Downloads and variants (7)

Every published repository of gpt-oss-20b-heretic-ara-v3, including quantised re-releases by other authors.

RepositoryFormatDownloads
mradermacher/gpt-oss-20b-heretic-ara-v3-GGUFGGUF4.7K
mradermacher/gpt-oss-20b-heretic-ara-v3-i1-GGUFGGUF (imatrix)1K
p-e-w/gpt-oss-20b-heretic-ara-v3sourceTransformers545
D8A/gpt-oss-20b-heretic-ara-v3-mlx-8BitMLX203
altomek/GPT-OSS-20B-Heretic-ARA-V3-GGUFGGUF68
Nograv/gpt-oss-20b-heretic-ara-v3Transformers6
dwedddd/gpt-oss-20b-heretic-ara-v3Transformers0

Security-team assessment

Derived from this model’s metadata and the publisher’s claims — not from benchmarks run by this site.

  • gpt-oss-20b-heretic-ara-v3 will attempt offensive-security prompts that a hosted commercial model declines, which is what makes it usable for red-team corpus generation and for measuring what an unaligned model of this class produces.
  • Capability is inherited from p-e-w/gpt-oss-20b-heretic-ara-v3, not added by ablation — at 20B parameters, expect roughly the base model’s competence, minus some instruction-following fidelity.
  • The publisher does not state which layers were ablated, so assume nothing about refusal consistency — probe it directly.
  • Declared licence is apache-2.0, but the terms that bind you are p-e-w/gpt-oss-20b-heretic-ara-v3’s — a re-release cannot grant more than its parent.

Handling: run it isolated, put moderation in front of it if anyone outside your team can reach it, and verify the weights before loading — these are third-party artifacts. Full handling and licence guidance.

Metadata for gpt-oss-20b-heretic-ara-v3 is collected from the public Hugging Face API and the publisher’s model card. This site does not host weights, is not affiliated with p-e-w, and does not independently verify publisher claims. See responsible use.