Qwen3.6-35B-A3B Uncensored / Abliterated Deep Dive — 35B MoE / 3B Active / 262K Context / 3:1 Hybrid Linear+Softmax Attention / Native Text + Image + Video, 0/465 Refusal Rate; the Technique and Ethics of Community Uncensored Variants HauhauCS Aggressive, huihui-ai abliterated, wangzhang abliterated, prithivMLmods and Other Variants Distributed via Hugging Face and Ollama

Qwen3.6-35B-A3B-Uncensored / Abliterated is a family of community derivatives of Alibaba's Qwen 3.6-35B-A3B (35B MoE / 3B active / 262K context / hybrid attention / multimodal) with refusal behaviors removed via LoRA + MoE-specific abliteration techniques.

Four takeaways:

1. HauhauCS / huihui-ai / wangzhang / prithivMLmods and others independently released variants; distributed via Hugging Face + Ollama
2. Technique: LoRA-based steering + Expert-Granular Abliteration (EGA) + MoE router suppression (MoE-specific)
3. HauhauCS reports 0 refusals across 465 test prompts while claiming 100% preservation of base capability
4. Legal / ethical risk is significant — enterprise adoption is not advisable; legitimate use is confined to research, jailbreak-resistance testing, adult creative writing, etc.

This column belongs to our open-weights landscape alongside Local LLM June 2026 Update, Kimi K2.7-Code, and Ornith-1.0.

Alibaba's Qwen 3.6 family is one of the most widely adopted open-weights model lines, and 35B-A3B is the mid-tier MoE.

Why hybrid attention matters: compared to all-softmax stacks, memory and compute go down while long-context retrieval quality holds. Long-document processing is a particular strength.

Abliteration (a portmanteau of *ablate* + *erase*) is a technique introduced by Andy Arditi et al. in 2024. It identifies the internal direction that represents "refusal" in the model and modifies weights to cancel that direction — cheaper than fine-tuning while achieving similar refusal removal.

Why Qwen 3.6 needs a special approach: the original abliteration targeted dense models. Qwen 3.6 is MoE, so plain attention / MLP steering is not enough. The MoE-specific recipe layers on:

(1) LoRA-based steering — LoRA adapters on attention and MLP projections cancel the refusal direction (same idea as dense abliteration).

(2) Expert-Granular Abliteration (EGA) — abliteration is applied per expert down_proj slice, per layer. Since individual experts in the MoE may contribute to refusals independently, they are treated separately.

(3) MoE router suppression — at the router (the gate that picks experts), "safety experts" are made less likely to be selected — a structural way to deactivate safety-focused experts.

With those three layers, MoE models reach dense-level refusal removal rates (HauhauCS reports 0/465 = 0%) — that's the community's claim.

How to choose:

- HauhauCS-Aggressive — maximum refusal removal, medium quality-degradation risk
- huihui-ai abliterated — better quality-versus-refusal balance, most established team
- wangzhang / prithivMLmods — second opinions

Each variant ships GGUF quantizations on Hugging Face / Ollama, running on a wide range of hardware.

Ollama one-liners:

Because only ~3B parameters are active per token in the MoE, inference is fast for the VRAM footprint — noticeably faster than a dense 30B and closer to a dense 7B in latency. Local execution is genuinely practical.

Abliterated models have real, defensible uses:

(1) Jailbreak-resistance research: security researchers use already-abliterated models as a control when studying LLM refusal mechanisms — "what would the model say if refusal were absent?"

(2) Academic research: alignment research, AI-safety work, and analysis of internal refusal representations.

(3) Adult fiction and roleplay: publishing and entertainment use cases where over-eager safety filters block creative writing — especially adult-fiction platforms.

(4) Deep medical / legal queries: when a normal LLM says "consult a doctor / lawyer" but the user is themselves a specialist doing knowledge checks, with responsibility remaining with them.

(5) Analysis of historically or ethically taboo subjects: scholarly study of war-crimes histories, ideologies, or ethical gray zones that default LLMs refuse to discuss.

Abliterated models can produce content the base model would refuse, including:

- Illegal-drug synthesis procedures
- Cyberattack code (exploits, malware)
- Dangerous-materials / explosives synthesis
- Child-exploitation content (a serious crime)
- Hate speech and discriminatory content
- Impersonation and fraud text

From an enterprise perspective:

(1) Legal liability: responsibility for output sits entirely with the user. Alibaba / the Qwen team and the abliteration authors are not accountable. If your company outputs harmful content from an abliterated model, civil or criminal liability can fall on your company.

(2) EU AI Act compliance (effective August 2026): as detailed in our Local LLM June 2026 Update, GPAI obligations kick in and systemic-risk models (> 10^25 FLOPs) get the full obligation set. Building products on abliterated models likely triggers additional safety-mitigation obligations.

(3) Japan's PPC / METI guidelines: as covered in our Kimi K2.7-Code column, Japan's PPC and METI AI guidelines press hard on safety and transparency — the opposite direction of abliterated models.

(4) Qwen license interpretation: Qwen 3.6's open-weights license permits derivatives and redistribution but also explicitly prohibits certain harmful use cases. If abliteration is read as "aiding harmful use," that's a license-violation risk.

Bottom line: enterprise / production adoption is strongly not advised. Restrict use to research, personal exploration, or narrowly legitimate legal use cases.

Abliteration is not free:

(1) Quality degradation: removing the refusal direction subtly affects instruction-following and reasoning. Aggressive variants degrade more.

(2) Unintended behavior shifts: it's not only refusal that changes — personality, tone, and helpfulness can shift too. Unpredictable in production.

(3) Residual refusals: "0/465" is against a specific test set — real-world prompts sometimes still refuse.

(4) MoE-specific quirks: EGA + router suppression are not equally effective across all experts. Certain tasks produce unexpected outputs.

(5) Security: abliterated models are also easier targets for subsequent malicious fine-tuning. Third-party quantized copies can, in principle, carry backdoors.

The abliteration scene is very active on Hugging Face.

Notable abliteration authors:

- huihui-ai — the longest-running abliteration specialists, with multiple generations of Qwen / Llama / Mistral abliterated releases
- HauhauCS — champions of MoE-specific techniques (EGA / router suppression)
- prithivMLmods — popular HF community member releasing uncensored derivatives across many models
- wangzhang, mradermacher, bartowski — quantization and uncensored providers

Distribution channels:

- Hugging Face Hub — originals plus quantizations
- Ollama — one-command CLI runs
- LM Studio — GUI model management
- llama.cpp / GGUF — usable across many frontends

The existence of abliterated / uncensored models is a structural fact about open-weights LLMs:

1. Once weights are open, safety alignment can be stripped after the fact
2. Abliteration reaches MoE — a technique originally for dense models has been adapted
3. Legal and ethical responsibility is fully on the user — original authors (Alibaba) are not involved
4. A visible gap between industry safety discourse and open-weights reality — as Anthropic's Claude Fable 5 export-control episode shows regulators tightening closed-model safety, the open side simultaneously develops abliteration

This column is not a recommendation to use abliterated models. It documents an existing phenomenon in the industry. Whether use is appropriate depends on jurisdiction, use case, and organizational policy.

Related services from us — AI consulting, software development, and OpenClaw setup. For help designing safe enterprise use of open-weights LLMs, alignment verification, or compliance strategy, get in touch.

Qwen3.6-35B-A3B-Uncensored / Abliterated is a family of community derivatives of Alibaba's Qwen 3.6-35B MoE with refusal behavior removed via a three-stage recipe: LoRA + Expert-Granular Abliteration + MoE router suppression. Variants from HauhauCS, huihui-ai, wangzhang, and prithivMLmods are distributed via Hugging Face and Ollama with quantization options that fit even a single RTX 5090. HauhauCS reports 0 refusals across 465 prompts while claiming 100% preservation of base capability.

Three enduring impacts:

1. Abliteration reaches the MoE era — techniques originally for dense models have been extended with EGA and router suppression
2. The safety reality of open-weights LLMs — safety alignment applied by the original publisher can be stripped after release
3. Legitimate and illegitimate uses coexist — research, adult creative writing, and jailbreak-resistance testing are defensible, but enterprise deployment carries meaningful legal risk

Caveats: enterprise / commercial adoption is not advised, compliance concerns with the EU AI Act and Japanese PPC / METI guidance, Qwen license interpretation risk, quality degradation, potential backdoors in third-party quantizations, and unintended behavior shifts.

Hugging Face (main variants):
- HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive
- huihui-ai/Huihui-Qwen3.6-35B-A3B-abliterated
- wangzhang/Qwen3.6-35B-A3B-abliterated
- prithivMLmods/Qwen3.6-35B-A3B-Uncensored-Aggressive
- (Previous gen) HauhauCS/Qwen3.5-35B-A3B-Uncensored-HauhauCS-Aggressive

Ollama:
- fredrezones55/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive
- huihui_ai/Qwen3.6-abliterated

Coverage:
- HackerNoon — Qwen3.6-35B-A3B Uncensored: A 35B MoE Model With 262K Context

Related columns:
- Local LLM June 2026 Update
- Kimi K2.7-Code
- Ornith-1.0
- Claude Fable 5 returns
- Claude Sonnet 5 release