Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-MLX-8bit

Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-MLX-8bit

MLX conversion of Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic for Apple silicon — 8-bit, measurably indistinguishable from bf16

Uncensored ("Heretic'd") multi-stage merge of Qwen3.5-9B fine tunes by nightmedia and DavidAU, with a compacted-but-stronger thinking block. Vision is included and works out of the box — no separate mmproj download.

Provenance

This was converted from nightmedia/Qwen3.5-9B-DS9-USS-Defiant (bfloat16 safetensors), which is the exact same weight set that DavidAU/Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-MTP-GGUF packages as GGUF.

We verified the identity numerically rather than assuming it:

So these MLX quants are made from the original bf16 weights, not by dequantizing a GGUF. There is no GGUF round-trip loss.

Credit for the model itself goes to nightmedia and DavidAU; this repo only does the MLX conversion.

Architecture

Qwen3.5-9B is a hybrid multimodal model:

• 32 layers, 3:1 ratio of gated-delta linear attention to full attention (full_attention_interval: 4)
• Gated output attention (attn_output_gate), head_dim 256, 16 Q heads / 4 KV heads
• Interleaved mRoPE with partial_rotary_factor 0.25, rope_theta 1e7
• 262,144 native context
• 27-layer vision tower (patch 16, spatial merge 2), 248,320 vocab

The source also ships MTP (multi-token-prediction) weights. MLX does not use them, so they are dropped during conversion — this costs no quality, only the speculative-decoding speedup that the GGUF "MTP" variants offer.

Usage

Vision + text with mlx-vlm:

pip install mlx-vlm
python -m mlx_vlm.generate \
  --model pipenetwork/Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-MLX-8bit \
  --image your_image.jpg \
  --prompt "Describe this image in detail." \
  --max-tokens 512

Text-only with mlx-lm (loads the same repo, ignores the vision tower):

pip install mlx-lm
mlx_lm.generate \
  --model pipenetwork/Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-MLX-8bit \
  --prompt "Write the opening paragraph of a noir story set on a space station." \
  --max-tokens 512

Python:

from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template

model, processor = load("pipenetwork/Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-MLX-8bit")
config = model.config

prompt = apply_chat_template(processor, config, "Describe this image.", num_images=1)
print(generate(model, processor, prompt, ["your_image.jpg"], max_tokens=512, verbose=False))

All quantizations, measured

Every tier below quantizes the identical bf16 weights, so bf16 is exact ground truth and the quantization scheme is the only variable. Measured on 65,536 tokens of wikitext-2 test at 1024 context, all models fed the same token ids through mlx-lm, on an M3 Ultra. KL is against bf16's own output distribution — lower means closer to the original model.

Sizes are the full repo; Bits/w is the whole-model average, which sits above the nominal width because the vision tower stays bf16 (0.91 GB) in every tier. All tiers are group size 64, affine mode.

Reading it:

• 8bit is effectively free — bf16 perplexity to four decimals at 47% of the footprint.
• 5bit is the best quality-per-GB: under 1% perplexity.
• The usable floor is 4bit. 3bit still answers factual questions correctly but costs +33% perplexity; treat it as a tight-memory fallback.
• No 2-bit tier is published. Pure 2-bit collapses (ppl 214.5, 28% top-1) and MLX's mixed_2_6 recipe, while grammatical, still runs ~7x bf16 perplexity at 3.94 GB — no smaller than 3bit and 5x worse. Both were built and measured; neither is usable.
• At the 4-bit tier this affine group-64 quant loses about half the perplexity MXFP4 does, costing 4.5 vs 4.25 bits/weight.

Reproduce with bench.py.

Sampling

DavidAU's notes for this model, which carry over:

• Temperature 1.0 or below works best; higher temps degrade coherence
• Repetition penalty 1.0 (off) — raising it hurts this model
• The thinking block is compacted; give it room with a generous max-tokens

Benchmarks

From the source model card (source weights, non-MLX quant tiers), for reference:

          arc/c  arc/e boolq hswag obkqa piqa  wino
bf16      0.649, 0.832, 0.895, 0.713, 0.482, 0.783, 0.699
mxfp8     0.647, 0.836, 0.895, 0.706, 0.460, 0.784, 0.695
mxfp4     0.640, 0.824, 0.886, 0.703, 0.468, 0.780, 0.691

Qwen3.5-9B-Instruct (base, non-heretic)
mxfp8     0.571, 0.719, 0.895, 0.683, 0.426, 0.770, 0.671

These are not measurements of these MLX repos — treat them as characterising the weights, not this quantization.

Conversion tooling

Scripts, benchmark and the provenance proof: github.com/PipeNetwork/defiant-fable-mlx

License

Apache 2.0, inherited from Qwen3.5-9B.

This model has had its safety post-training removed and will follow instructions without refusal. You are responsible for how you use it.