library_name: mlx-vlm
pipeline_tag: image-text-to-text
base_model: huihui-ai/Huihui-Qwen3.6-27B-abliterated
base_model_relation: quantized
tags:
- mlx
- mlx-vlm
- qwen3.6
- multimodal
- vision-language
- abliterated
Huihui-Qwen3.6-27B-abliterated-mlx-mxfp8
Local MLX-VLM conversion of huihui-ai/Huihui-Qwen3.6-27B-abliterated.
Overview
- Format: MLX-VLM
- Precision: mxfp8
- Size: 27G
- q_mode: mxfp8
- q_bits: 8
- Source model type: Qwen3_5ForConditionalGeneration
- Source pipeline: image-text-to-text
- Intended runtime: mlx-vlm, LM Studio
Validation
- text generation smoke test: passed
- image generation path smoke test: passed
Quantization
- Quantization mode: mxfp8
- Quantization bits: 8
Test Results (2026-04-23)
Test runner: mlx_vlm.generate
Text prompt:你好，请用一句自然中文回应。
Image prompt:请用一句中文描述这张图片。
Image asset:
64x64 pure red square
Observed results:
| Case | Prompt TPS | Generation TPS | Peak Memory |
|---|---|---|---|
| Text | 32.519 | 11.425 | 28.810 GB |
| Image | 84.449 | 11.386 | 28.980 GB |
Notes:
- Both text and image tests completed successfully.
- Compared with bf16, this build materially reduces peak memory while preserving the basic text and image inference path.
- The current
mlx_vlmruntime still emits<think>content even with--processor-kwargs '{"enable_thinking": false}'.