This repository contains an oMLX oQ8 mixed-precision MLX quantization of llmfan46/Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved with both the vision encoder and Native MTP tensors preserved.

This is a replacement build for the earlier Native-MTP-Preserved MLX artifact that preserved MTP but omitted the vision-side tensors. The source VLM config is kept functionally equivalent, with vision_config.model_type patched to qwen3_5 for current MLX-VLM/oMLX compatibility.

Variant

Quantization: oQ8
Variant: Native MTP Preserved VLM
Vision tensors: preserved
MTP tensors: preserved
Text-only: false
Approximate target density: 8.63 bpw.

Usage

omlx serve dawncr0w/Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved-oQ8-MLX

For direct Python loading, use the VLM loader from the bundled oMLX/MLX-VLM runtime.

Validation

Local validation completed with the bundled oMLX runtime:

text

loader: mlx_vlm.load_model with oMLX VLM/MTP patches
load smoke test: passed
peak memory: 27.942 GB
visual tensors: 333
mtp tensors: 29

Source

BF16 source checkpoint: llmfan46/Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved
BF16 GGUF reference: llmfan46/Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved-GGUF
Quantization tool: oMLX oQ

Upstream model card license tag: apache-2.0.