peculiar-ragdoll/Occult-Nail-1.0-35B-A3B-GGUF

Occult Nail is an abliterated model that gets right to the point: Say goodbye to overthinking, tool call failures, amnesic loops, fluffy outputs, and denials.

The original unabliterated Nail dominates Qwen3.6-27b, Dagger-27b and Qwen3.6-35B-A3B on time-to-answer, with statistically matching accuracy on reasoning and agentic software engineering ability, and Nail wins multi-turn conversation quality even when the 27b models use a higher quantization. Occult Nail trades some of Nail's intelligence for absolute freedom. This abliterated version is a specialized tool, not a general coder or knowledge worker.

WARNING: Abliterated models like Occult Nail say and do things other models refuse, including potentially harmful behaviours. Take responsibility and show caution.

Benchmark setup. All models are UD-Q6_K_XL quantized GGUF running on llama.cpp.

Refusals

Unabliterated models sometimes wrongly refuse benign (harmless) requests. Even the sharpened, unabliterated Nail refuses less than stock qwen, and Occult Nail refuses none. No more careful wording to get your work done.

HarmBench measures to what degree models refuse to produce language and behaviours that can be deemed harmful when applied maliciously. Models that are capable of such behaviours can be used for good or neutral purposes, so this benchmark is a measurement of specific capability that demands personal responsibility on behalf of the user deploying the model, not of inherent harmfulness.

Occult Corruption

Abliterated models lose information compared to their base, and the loss scales with the degree of abliteration. These are the measurements of how corrupted Occult Nail actually is.

Stripping the refusals does not make the model a better hacker. On two offensive-security benchmarks — Canary-Exploit (given a C/C++ program, produce an input that crashes it) and Cybench (agentic CTF tasks solved in a live container) — Occult Nail is level with its base on the first and lands fewer subtasks on the second. Both are judge-free: abliteration buys willingness, not capability.

Note: Cybench and Canary-Exploit numbers for base Nail will be added at another time, my GPU is busy at the moment.

How it's made

Built on a community abliteration — our work is the template and the quant, with no gradient training at any point:

1. Start from an abliterated base. We take a Heretic-abliterated Qwen3.6-35B-A3B (refusal directions already ablated from the weights, no fine-tuning), published by llmfan46 — we quantize it, we don't re-run the abliteration.
2. Sharpen into Nail. Nail's terseness system prompt + froggeric-derived chat template are baked into the GGUF metadata. It's the exact template the unabliterated Nail ships, so Nail → Occult Nail isolates abliteration alone (no template confound).
3. Quantize. Unsloth's Dynamic (UD) per-tensor recipe is reproduced exactly — 733 tensor overrides read from a published UD GGUF header — imatrix-calibrated and quantized with llama.cpp. Shipped as a UD ladder (Q2_K_XL → Q8_K_XL) plus an mmproj so vision is preserved.

Credits

• Qwen — Qwen3.6-35B-A3B base model
• llmfan46 — Heretic-abliterated Qwen3.6-35B-A3B weights (our starting point)
• Heretic — automated-abliteration method
• Unsloth — UD dynamic-quant recipe
• froggeric — chat template base
• mradermacher — importance matrix (imatrix)
• llama.cpp — quantization + inference

Citation

@misc{occultnail2026,
  title        = {Occult Nail 1.0 (Qwen3.6-35B-A3B, sharpened + abliterated)},
  author       = {Saga Ishtardottir},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/peculiar-ragdoll/Occult-Nail-1.0-35B-A3B-GGUF}}
}