⌦abliteratedmodels.org

K2-Horizon-MoVA-36B-A4B-heretic

K2-Horizon-MoVA-36B-A4B-heretic is a Heretic-ablated variant of darkc0de/K2-Horizon-MoVA-36B-A4B-heretic, published on Hugging Face by darkc0de. It is 36B total parameters (mixture-of-experts), 48GB VRAM class class and apache-2.0 licence. It is distributed across 4 repositories in Transformers, GGUF and GGUF (imatrix) formats, totalling 11K downloads.

Heretic36B MoEOther25B – 50B↓ 10,971♥ 3
View on Hugging Face ↗Base model ↗

Specification

Model name
K2-Horizon-MoVA-36B-A4B-heretic
Source repository
darkc0de/K2-Horizon-MoVA-36B-A4B-heretic
Base model
darkc0de/K2-Horizon-MoVA-36B-A4B-heretic
Publisher
darkc0de
Parameters
36B (mixture-of-experts)
VRAM required
~22.7 GB at Q4_K_M, 4k context (estimated) — runs on 24 GB cards
Hardware class
25B – 50B — 48GB VRAM class
Licence
apache-2.0
Task
Text generation
Context window
512K tokens (as documented)
Ablation scope
Not stated by the publisher
Formats available
Transformers, GGUF, GGUF (imatrix)
First indexed
17 Sept 2026
Last updated
14 Oct 2026

Ablation technique

How K2-Horizon-MoVA-36B-A4B-heretic 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

“K2-Horizon-MoVA-36B-A4B is the sparse member of the K2-Horizon family: a Mixture-of-Experts model with Mixture-of-Values attention (MoVA) that stores 36B parameters and runs 4B per token. We have released the final checkpoint; intermediate checkpoints, along with the data and the training code, will be released.”

Running it

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

llama.cpp / GGUF
hf download mradermacher/K2-Horizon-MoVA-36B-A4B-heretic-i1-GGUF --local-dir ./k2-horizon-mova-36b-a4b-heretic
llama-cli -m ./k2-horizon-mova-36b-a4b-heretic/<file>.gguf -p "..."
Transformers
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "darkc0de/K2-Horizon-MoVA-36B-A4B-heretic"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(
    repo, torch_dtype="auto", device_map="auto"
)
vLLM
vllm serve darkc0de/K2-Horizon-MoVA-36B-A4B-heretic --trust-remote-code

Downloads and variants (3)

Every published repository of K2-Horizon-MoVA-36B-A4B-heretic, including quantised re-releases by other authors.

RepositoryFormatDownloads
mradermacher/K2-Horizon-MoVA-36B-A4B-heretic-i1-GGUFGGUF (imatrix)9.9K
mradermacher/K2-Horizon-MoVA-36B-A4B-heretic-GGUFGGUF911
darkc0de/K2-Horizon-MoVA-36B-A4B-hereticsourceTransformers177

Security-team assessment

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

  • K2-Horizon-MoVA-36B-A4B-heretic 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 darkc0de/K2-Horizon-MoVA-36B-A4B-heretic, not added by ablation — at 36B 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 darkc0de/K2-Horizon-MoVA-36B-A4B-heretic’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 K2-Horizon-MoVA-36B-A4B-heretic is collected from the public Hugging Face API and the publisher’s model card. This site does not host weights, is not affiliated with darkc0de, and does not independently verify publisher claims. See responsible use.