Method: OptiQ data-driven mixed precision (optiq_mixed_precision)
Average: ~5.6 bits per weight (220 layers at 8-bit, 276 at 4-bit)
Architecture: qwen3_5 (Qwen3.6-27B)
Reasoning: this is a thinking model; append :think / :no-think to the model id, or set enable_thinking in the chat template, to control the thinking channel

The per-layer allocation is reused from mlx-community/Qwen3.6-27B-OptiQ-4bit, which shares the same Qwen3.6-27B architecture, so no separate sensitivity pass was run for this fine-tune. This is a text-only quant of the language tower.

Use

bash

pip install mlx-optiq
optiq serve --model mlx-community/Qwen3.6-27B-Heretic2-Uncensored-Finetune-Thinking-OptiQ-4bit

Or load it directly with mlx_lm:

python

from mlx_lm import load, generate
model, tokenizer = load("mlx-community/Qwen3.6-27B-Heretic2-Uncensored-Finetune-Thinking-OptiQ-4bit")

See mlx-optiq.com for the CLI, the local Lab UI, and the OptiQ Code agent. The base model and its behavior are DavidAU's; this repo only changes the quantization.