iamthecage/Qwen3.6-27B-Thinking-MLX-mixed-7.6bit-MTP

<div align="center"><img width="400px" src="https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3.6/logo.png"></div>

Qwen3.6-27B-Heretic2-Uncensored-Finetune-Thinking

*Added full MTP support, KV Quantization Merged jinja chat template fix from froggeric/Qwen-Fixed-Chat-Templates*

Instructions

You will need this branch of mlx-lm:

https://github.com/iamthecage/mlx-lm

Install mlx-lm from the cloned repo:

pip install -e . --no-build-isolation

Optional(highly recommended): Install OptiQ to enable KV Quantization and fused SPDA optimizations

pip install mlx-optiq

Serve with:

mlx_lm.server \                                 
  --model iamthecage/Qwen3.6-27B-Thinking-MLX-mixed-7.6bit-MTP \
  --mtp-draft --num-draft-tokens 4 \
  --kv-bits 8 --kv-group-size 64 --quantized-kv-start 2048 \
  --chat-template-args '{"enable_thinking": true, "preserve_thinking": true, "auto_disable_thinking_with_tools": true}' \
  --temp 0 --host 127.0.0.1 --port 8082

MTP Enabled INFO - Metrics: prompt 23123 tok @ 318.8 tok/s | gen 157 tok @ 20.2 tok/s | peak 34.22 GB | mtp 116/157 accepted (74%) accept-len 3.83

MTP Disabled INFO - Metrics: prompt 28152 tok @ 362.8 tok/s | gen 144 tok @ 9.8 tok/s | peak 34.22 GB

Quality: quantized (*mixed quants per tensor, group size: 32, 7.652 bpw*)

Most tensors use 6-bit or 8-bit affine quantization with a group size 32.

Fully uncensored and then fine-tuned by DavidAU.

Abliteration metrics

IN HOUSE BENCHMARKS [by Nightmedia]:
         arc-c arc/e boolq hswag obkqa piqa  wino
Qwen3.6-27B-Heretic2-Uncensored-Finetune-Thinking
mxfp8    0.673,0.846,0.905... [instruct mode]
Qwen3.6-27B-Heretic-Uncensored-Finetune-Thinking
mxfp8    0.669,0.835,0.906,...   [instruct mode]
---
BASE UNTUNED MODEL:
Qwen3.6-27B HERETIC (by llmfan46) [instruct mode]
mxfp8    0.644,0.788,0.902,...
Qwen3.6-27B (by Qwen) [instruct mode]
mxfp8    0.647,0.803,0.910,0.773,0.450,0.806,0.742

Source

This model was converted to MLX format from `DavidAU/Qwen3.6-27B-Heretic2-Uncensored-Finetune-Thinking` using mlx-vlm version 0.4.4.