Model details
- Base model: Qwen/Qwen3.6-35B-A3B (Apache-2.0)
- Architecture:
Qwen3_5MoeForConditionalGeneration (qwen3_5_moe) · MoE ~34.7B total / ~3B active
- Method: Heretic directional-refusal ablation — 200 optimization trials, low post-ablation KL divergence (≈ 0.0003 vs. base)
- Format: bf16
safetensors (16 shards, ~66 GB) · 131K context
Running it
qwen3_5_moe is supported by recent llama.cpp (master / ~b6900+), so GGUF now works — alongside vLLM and MLX:
- GGUF / llama.cpp (any platform, incl. Apple Metal):
…-gguf — Q6_K (~27 GB) / Q4_K_M (~20 GB). Needs a build new enough to include the qwen3_5_moe arch.
- NVIDIA → vLLM. fp8 fits a single 46 GB L40S; native tool-calling via
--tool-call-parser qwen3_xml.
- Apple Silicon → MLX 4-bit:
…-mlx-4bit.
Usage (vLLM)
vllm serve jbrahy/Qwen3.6-35B-A3B-Never-Never-LAN-uncensored-abliterated \
--quantization fp8 --language-model-only --max-model-len 131072 \
--enable-auto-tool-choice --tool-call-parser qwen3_xml \
--host 0.0.0.0 --port 8083
Limitations & responsible use
Abliteration removes safety refusals — this model will answer prompts the base model would decline. You are responsible for how you use it. It inherits the base model's biases, knowledge cutoff, and capabilities. Licensed Apache-2.0, same as the base.
