license: apache-2.0
language:
- en
- zh
tags: - mlx
- qwen3.5
- vlm
- vision
- abliterated
- uncensored
pipeline_tag: image-text-to-text
Qwen3.5 4B Claude Opus Abliterated — MLX VLM
Abliterated Qwen3.5 4B vision-language model, converted for mlx-swift-lm compatibility.
About
This model is derived from mlx-community/Huihui-Qwen3.5-4B-Claude-4.6-Opus-abliterated-6bit.
The original model works with mlx-vlm but had issues with mlx-swift-lm:
- ❌ Text-only chat template (no image support)
- ❌ Nested processor_config.json format
What was fixed
| Issue | Fix |
|---|---|
| chat_template.jinja | Replaced with VLM-aware template handling image/video content arrays |
| processor_config.json | Added flat-format preprocessor_config.json for Qwen3VLProcessor |
Usage with mlx-swift-lm
import MLXVLM
import MLXLMCommon
import Tokenizers
let modelDir = URL(fileURLWithPath: "/path/to/model")
let container = try await VLMModelFactory.shared.loadContainer(
from: modelDir,
using: LocalTokenizerLoader()
)
let session = ChatSession(container)
let response = try await session.respond(
to: "Describe the image",
image: .url(URL(fileURLWithPath: "/path/to/image.jpg"))
)
Download
To get the complete model, download the weight files from the original repo, then replace the config files with the fixed versions from this repo.
Fixed by monyschuk/mlx-qwen-fix