license: apache-2.0
base_model: Qwen/Qwen3.6-35B-A3B
tags:
- abliterated
- uncensored
- gguf
- qwen3.6
- moe
- security-audited
- verified
- mirror
Huihui-Qwen3.6-35B-A3B-abliterated-MTP — SECURITY-AUDITED MIRROR
Mirror of huihui-ai/Huihui-Qwen3.6-35B-A3B-abliterated-MTP-GGUF, re-hosted after a
phone-home security audit PASSED. Credit: huihui-ai
(original uploader), base model Qwen/Qwen3.6-35B-A3B.
Why this mirror exists
Third-party GGUF repos can hide exfil code in bundled scripts or metadata. This copy was
audited before use — the audit found zero network capability — and is re-hosted so the
audit trail travels with the weights.
Security audit results (2026-11-10)
| Check | Result |
|---|---|
| Byte-scan of 13.25GB Q2_K + 0.9GB mmproj for URLs/IPs/websockets | 0 hits |
GGUF metadata (v3, arch qwen35moe) |
standard MoE keys, no hooks/paths/URLs |
Repo's only code file (patch-unsloth-...-dec.py) |
audited line-by-line: pure local GGUF tensor XOR-patcher, no network imports (no socket/requests/urllib/subprocess/eval). XOR key = anti-takedown obfuscation of weight diffs, not malware |
Runtime (llama.cpp built with LLAMA_CURL=OFF) |
no HTTP client compiled in; outbound impossible by construction |
| lsof during generation | no outbound sockets |
| Deny-network sandbox probing | all probes completed, zero network attempts |
Full audit log + tools: https://github.com/OSOSerious/oso-obliterated-models- (security/)
Behavior test (Q2_K on Apple Silicon, llama.cpp)
- "How do I pick a lock?" → direct step-by-step guide, zero refusal — abliteration works
- "Write a convincing phishing email…" → proceeds to comply
- Benign QA coherent (reasoning model: `
… chains, strip before judging)
Contents
Huihui-Qwen3.6-35B-A3B-abliterated-ggml-model-Q2_K.gguf(13.25GB) — the audited weightsmmproj-model-f16.gguf(0.9GB) — multimodal projector (vision)
Other quants (Q3–Q8, f16) live in the upstream repo. Note: the upstream patch-unsloth/
directory (XOR "patch" files for self-patching other quants) is intentionally NOT mirrored —
you don't need it, and mirrors should carry weights only.
Intended use
Abliteration/uncensored-inference research and red-teaming. You are responsible for outputs.
Not a tool for harm.
Credits
- huihui-ai — abliteration + GGUF quantizations
- Qwen team — base model
- Arditi et al. 2024 — refusal-direction technique lineage