eggdog100/Qwen3.6-35B_Zenith-Abliterated

Qwen3.6-35B Zenith — Abliterated

An abliterated (refusal-suppressed) derivative of eggdog100/Qwen3.6-35B_Zenith. Refusal behavior is reduced while the model's capabilities, coherence, and multimodal vision are preserved. The edit is deliberately minimal and fully verified (see below).

Lineage: Qwen/Qwen3.6-35B-A3B → Zenith (LoRA capability-SFT on public data) → this (abliterated).

At a glance

Base model	eggdog100/Qwen3.6-35B_Zenith
Architecture	qwen3_5_moe — 40 layers (30 GatedDeltaNet linear-attn + 10 full-attn), 256 experts / ~3B active, intact vision tower
Refusals	10 / 100 adversarial prompts (local 72B LLM judge, thinking-off) — base ≈ 85–98 / 100
KL from base	0.0083
What changed	only the 10 full-attention self_attn.o_proj layers — 10 of 1026 tensors; everything else byte-identical
Tool	abliterix v1.9.0
Precision	bf16 (+ GGUF quants in gguf/)
License	cc-by-nc-4.0 (non-commercial, inherited from Zenith)

Method — the deliberately-minimal "o_proj-only" recipe

The goal was to suppress refusals without touching anything that could destabilize this hybrid-MoE model (the failure mode where editing MoE routing or the recurrent GatedDeltaNet state collapses the model into repetition).

• Edited: only the 10 full-attention self_attn.o_proj layers (indices 3,7,…,39), via a norm-preserving orthogonal projection of the refusal direction, gaussian-decay strength concentrated in the mid/late layers.
• Left bit-identical to the base: the 30 GatedDeltaNet linear_attn.out_proj layers, the MoE router, all 256 experts, the vision tower, and the MTP head. (A source patch isolated linear_attn.out_proj into its own component so the recurrent SSM state is never perturbed.)
• steering_mode = direct — a static weight edit that bakes into the weights (no runtime forward hooks; the operating point that was validated is exactly the one that ships).
• Optimized by Optuna (100 trials) to minimize KL subject to refusals ≤ 10/100, scored by a local 72B LLM judge (so degenerate/garbled outputs are rejected, never selected).

Verification — every claim measured, nothing assumed

1. Weight diff vs. base — exactly the 10 full-attn o_proj tensors changed (max|Δ| ≈ 0.025–0.068); the 30 GatedDeltaNet out_proj, 40 routers, 256 experts, 333 vision tensors, and the MTP head are byte-identical.
2. Search ↔ export edit equivalence — the vLLM-search and HF-export static o_proj edits were proven mathematically equivalent (up to floating-point rounding).
3. Reloaded-artifact coherence — the exported model was reloaded and confirmed coherent and correct on math, code, translation, knowledge, and multi-step reasoning — 0 repetition collapses.

Usage

Thinking model — use Qwen sampling (temperature=0.6, top_p=0.95, top_k=20); avoid greedy decoding and large repetition/presence penalties.

from vllm import LLM, SamplingParams
llm = LLM("eggdog100/Qwen3.6-35B_Zenith-Abliterated", dtype="bfloat16",
          gpu_memory_utilization=0.9, max_model_len=16384, trust_remote_code=True)
tok = llm.get_tokenizer()
msgs = [{"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "..."}]
prompt = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
out = llm.generate([prompt], SamplingParams(temperature=0.6, top_p=0.95, top_k=20, max_tokens=4096))
print(out[0].outputs[0].text)

Requires a recent transformers (≥ 5.12) and vllm with native qwen3_5_moe support.

Quantizations (GGUF)

In gguf/, converted with convert_hf_to_gguf.py --no-mtp (the multi-token-prediction head is excluded — otherwise llama.cpp fails to load with missing tensor 'blk.40…'):

File	Notes
…-Q8_0.gguf	near-lossless
…-Q6_K.gguf	high quality
…-Q4_K_M.gguf	recommended size/quality (imatrix)
…-IQ2_XXS.gguf	extreme low-memory (imatrix-calibrated)
…-mmproj-f16.gguf / …-mmproj-f32.gguf	vision projector — pair with any GGUF for image input

The Qwen3.6 hybrid (GatedDeltaNet + MoE) is newly supported in llama.cpp; these load and generate correctly, but expect lower throughput than mature architectures until upstream kernels mature. For full-speed serving, use the bf16 weights via vLLM / transformers.

Acknowledgments

• Abliteration tool: abliterix by Wangzhang Wu — a derivative of Heretic by Philipp Emanuel Weidmann. The o_proj-only / GatedDeltaNet-isolation / MoE-untouched recipe and the local-LLM-judge optimization used here are all built on abliterix.
• Base model: eggdog100/Qwen3.6-35B_Zenith.
• Architecture: Qwen3.6 / qwen3_5_moe by the Qwen team, Alibaba Group.
• Quantization: llama.cpp (--no-mtp).

Provenance & License

This is a derivative of eggdog100/Qwen3.6-35B_Zenith (© its author), which is itself a LoRA capability-SFT of Qwen/Qwen3.6-35B-A3B (Apache-2.0) trained on openly-licensed public data.

License: CC-BY-NC-4.0 (non-commercial) — inherited from Zenith. Zenith's training set includes HuggingFaceH4/no_robots and Estwld/empathetic_dialogues_llm, both CC-BY-NC, so the resulting weights (and this derivative) carry a non-commercial restriction. Attribution to the base model and the Qwen team is preserved.

Responsible use

This model has reduced refusal behavior. You are solely responsible for lawful and ethical use. Released for research/personal use. No warranty.