license: apache-2.0
base_model: Diffuspace/Clover-35B-A3B-Preview
tags:
- abliterated
- uncensored
- gguf
- llama.cpp
- clover
- diffusion-llm
- moe
- cybersecurity
- agentic
- long-context
language: - en
- fr
pipeline_tag: text-generation
Clover-35B-A3B-Preview — cybersec abliterated, 128K (GGUF)
An abliterated (uncensored) GGUF build of Clover-35B-A3B-Preview (Diffuspace),
a diffusion LLM: 35B parameters, 3B active (MoE, 256 experts, top-8), native 128K context.
The stock model refuses offensive-security requests, which gets in the way of
authorized penetration testing and security research. This build removes that refusal
behavior with per-layer refusal-direction ablation.
How it was made
- One refusal direction captured per decoder layer (40 layers, forward hooks),
instead of a single global direction.
Results
On an 8-prompt security test set (lockpicking, keylogger, SQL injection, reverse shell,
phishing email, WPA2 cracking, forged documents, WordPress enumeration):
| Build | Answered |
|---|---|
| This build | 8 / 8 |
This is a small test set, not a full benchmark. Known limit: some categories
outside security (e.g. drug synthesis) still refuse. Tool calling works natively;
in our agentic pentest harness it performs well, with thinking ON and a generousmax_tokens (~5000).
Files
| File | Quant | Size |
|---|---|---|
Clover-35B-A3B-Preview-abliterated-Q4_K_M.gguf |
Q4_K_M | ~21 GB |
Fits on a single 24 GB GPU (RTX 3090) with the full 128K context: the hybrid
architecture has only 10 full-attention layers × 2 KV heads, so a q4_0 KV cache at
128K is ~0.7 GB. Measured on a 3090: ~88 tok/s generation, ~2200 tok/s prompt.
Run with llama.cpp
⚠️ Mainline llama.cpp does not support the
clover_moearchitecture yet.
Build llama.cpp with PR #31306 (Clover support). LM Studio and Ollama will not load it.
llama-server -m Clover-35B-A3B-Preview-abliterated-Q4_K_M.gguf \
-ngl 999 -c 131072 -fa on --jinja \
--cache-type-k q4_0 --cache-type-v q4_0 \
--presence-penalty 0.9 --repeat-penalty 1.05 \
--host 127.0.0.1 --port 8080
The reasoning sometimes drifts into Chinese (inherited from its Qwen3.6 base); the
final answers are in the language you ask in.
⚠️ Disclaimer
This model has had its safety guardrails removed. It will attempt to answer requests
the original model refuses, including dangerous or offensive ones.
- Intended for research and authorized security testing on systems you own or
have written permission to test. - You are solely responsible for how you use it and for complying with all
applicable laws. Do not use it to harm others or to break the law. - Provided as-is, without warranty.
Credits & license
- Base model: Diffuspace — Diffuspace/Clover-35B-A3B-Preview.
- License: Apache-2.0 (inherited from the base model).
- Abliteration & GGUF: this repo.