Instructions to use pipenetwork/Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-MLX-4bit with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use pipenetwork/Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-MLX-4bit with MLX:

# Make sure mlx-vlm is installed
# pip install --upgrade mlx-vlm

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

# Load the model
model, processor = load("pipenetwork/Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-MLX-4bit")
config = load_config("pipenetwork/Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-MLX-4bit")

# Prepare input
image = ["http://images.cocodataset.org/val2017/000000039769.jpg"]
prompt = "Describe this image."

# Apply chat template
formatted_prompt = apply_chat_template(
    processor, config, prompt, num_images=1
)

# Generate output
output = generate(model, processor, formatted_prompt, image)
print(output)

Notebooks
Google Colab
Kaggle
Local Apps Settings
LM Studio

How to use pipenetwork/Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-MLX-4bit with Pi:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "pipenetwork/Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-MLX-4bit"

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "mlx-lm": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "pipenetwork/Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-MLX-4bit"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use pipenetwork/Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-MLX-4bit with Hermes Agent:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "pipenetwork/Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-MLX-4bit"

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default pipenetwork/Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-MLX-4bit

Run Hermes

hermes

OpenClaw new

How to use pipenetwork/Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-MLX-4bit with OpenClaw:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "pipenetwork/Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-MLX-4bit"

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "pipenetwork/Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-MLX-4bit" \
  --custom-provider-id mlx-lm \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-MLX-4bit / README.md

pipenetwork

Add 5bit tier to the measured quantization comparison

15f8b33 verified 8 days ago

7.29 kB

language:
  - en
  - zh
license: apache-2.0
library_name: mlx
pipeline_tag: image-text-to-text
tags:
  - mlx
  - apple-silicon
  - qwen3.5
  - fine tune
  - heretic
  - uncensored
  - abliterated
  - merge
  - thinking
  - reasoning
  - creative
  - writing
  - fiction
  - roleplaying
  - vision
base_model: nightmedia/Qwen3.5-9B-DS9-USS-Defiant

Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-MLX-4bit

MLX conversion of Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic for Apple silicon — 4-bit, the smallest tier

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:

Check	Result
`lm_head.weight` (BF16 in both) vs GGUF `output.weight`	bit-exact, 0 mismatches across 262,144 values
RMSNorm weights (`input_layernorm`, `q_norm`, `post_attention_layernorm`, final `norm`)	equal `source + 1.0` under llama.cpp's shift convention (12542/12544 elements exact; 2 differ by 1 ULP from the float32 addition)
`embed_tokens` vs GGUF Q8_0 `token_embd`	cosine 0.999956 — consistent with a plain Q8_0 round-trip

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-4bit \
  --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-4bit \
  --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-4bit")
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))

Quantizations

Repo	Bits	Size
...-MLX-4bit	4	6.0 GB
...-MLX-6bit	6	7.7 GB
...-MLX-5bit	5	6.6 GB
...-MLX-8bit	8	9.7 GB
...-MLX-bf16	16	18.8 GB

Group size 64, affine mode. The vision tower is left unquantized (mlx-vlm's default for multimodal projector/patch-embed modules), so the size delta between tiers comes from the language model.

Measured quantization quality

Against the bf16 reference these were converted from — 65,536 tokens of wikitext-2 test at 1024 context, identical token ids through mlx-lm, on an M3 Ultra. KL is measured against bf16's own output distribution, so lower means closer to the original model.

model	lang. weights	ppl	Δppl	KL(bf16‖quant)	top-1 vs bf16	decode
bf16 reference	17.91 GB	8.1273	—	—	—	39.2 t/s
8bit	9.51 GB	8.1277	+0.00%	0.00124	98.24%	67.9 t/s
6bit	7.28 GB	8.1426	+0.19%	0.00523	96.20%	80.9 t/s
5bit	6.16 GB	8.2012	+0.91%	0.01845	93.24%	92.5 t/s
4bit (this repo)	5.04 GB	8.5816	+5.59%	0.07330	87.06%	110.0 t/s
nightmedia mxfp4	4.76 GB	8.9443	+10.05%	0.11328	82.34%	114.7 t/s

Reading it:

8bit is effectively free — bf16 perplexity to four decimals at 47% of the footprint and 1.7x the decode speed.
5bit is the best quality-per-GB: under 1% perplexity for 6.2 GB.
The cliff is 5bit → 4bit, where KL jumps 4x and perplexity goes +0.91% → +5.59%. If 4bit feels lossy, 5bit is the tier to move to, not 6bit.
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 (~6% more storage, ~4% slower decode).

All tiers are quantizations of the same weights, so this isolates the quantization scheme. Reproduce with bench.py.

Sampling## 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.