language:

en
fi license: apache-2.0 base_model:
Qwen/Qwen3.6-27B tags:
mlx
oQ4
omlx
quantization
qwen
qwen3.6
fable
fusion
heretic
uncensored
abliterated
DavidAU
MTP GGUF Quants
creative writing
roleplaying
fiction pipeline_tag: text-generation library_name: mlx

Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU — oQ4 MLX Quant

oMLX oQ4 recipe-driven quantization of the DavidAU Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU merge (Qwen3.6-27B base). Per-tensor hard floors preserved on all 298 recipe-pinned tensors; only the base tier was downgraded to oQ4 to land a smaller file.

Quantization summary


Quant method	oMLX oQ, enhanced
Base level	`oQ4` — base bits 4, group size 64, affine
Recipe pins	298 tensors (187 × 8-bit, 111 × 6-bit) — preserved verbatim
oQ4 self-boosts	45 extra tensors lifted to 5-bit (linear-attention, calibration-driven)
Final override map	343 tensors (187 × 8-bit, 111 × 6-bit, 45 × 5-bit)
Model size	20.16 GB on disk (5 shards)
Language model	~21.6 GB dry-run estimate at oQ6e reference; actual oQ4 build ~19.7 GB LM
Text-only	vision encoder stripped
MTP	stripped
Dtype	float16

Recipe / pinned floors

FA6 — 6-bit floor on all 16 full-attention layers (self_attn + MLP)
QKV6 — 6-bit floor on all linear_attn.in_proj_qkv
Tail6 — 6-bit floor on layers 56–58
E6 — embed_tokens at 6-bit
b8 — 8-bit bump set (187 tensors, incl. lm_head)

Verified: recipe → config override map matches exactly (0 missing, 0 bit mismatches); on-disk safetensors dtypes confirm pinned tensors are packed-quantized (U32), not fp16.

Use with mlx

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-oQ4-fa6-qkv6-tail6-e6-b8")

prompt = "hello"

if tokenizer.chat_template is not None:
    messages = [{"role": "user", "content": prompt}]
    prompt = tokenizer.apply_chat_template(
        messages, add_generation_prompt=True, return_dict=False,
    )

response = generate(model, tokenizer, prompt=prompt, verbose=True)

About the base

This MLX quant is derived from the fp source Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU (Qwen3.6-27B base, Fable-Fusion-711 merge family, uncensored / abliterated, DavidAU). 64-layer hybrid architecture: 48 linear-attention + 16 full-attention layers. Original model license: apache-2.0.

Quantization summary

Quant method

oMLX oQ, enhanced

Base level

oQ4 — base bits 4, group size 64, affine

Recipe pins

298 tensors (187 × 8-bit, 111 × 6-bit) — preserved verbatim

oQ4 self-boosts

45 extra tensors lifted to 5-bit (linear-attention, calibration-driven)

Final override map

343 tensors (187 × 8-bit, 111 × 6-bit, 45 × 5-bit)

Model size

20.16 GB on disk (5 shards)

Language model

~21.6 GB dry-run estimate at oQ6e reference; actual oQ4 build ~19.7 GB LM

Text-only

vision encoder stripped

MTP

stripped

Dtype

float16

Recipe / pinned floors

FA6 — 6-bit floor on all 16 full-attention layers (self_attn + MLP)

QKV6 — 6-bit floor on all linear_attn.in_proj_qkv

Tail6 — 6-bit floor on layers 56–58

E6 — embed_tokens at 6-bit

b8 — 8-bit bump set (187 tensors, incl. lm_head)

Verified: recipe → config override map matches exactly (0 missing, 0 bit mismatches); on-disk safetensors dtypes confirm pinned tensors are packed-quantized (U32), not fp16.

Use with mlx

pip install mlx-lm

from mlx_lm import load, generate model, tokenizer = load("DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-oQ4-fa6-qkv6-tail6-e6-b8") prompt = "hello" if tokenizer.chat_template is not None: messages = [{"role": "user", "content": prompt}] prompt = tokenizer.apply_chat_template( messages, add_generation_prompt=True, return_dict=False, ) response = generate(model, tokenizer, prompt=prompt, verbose=True)