v102ss/Alice-Qwen3.5-4B-Heretic-Light-MLX-4bit

Alice Qwen3.5 4B Heretic Light MLX 4bit

This is the 4bit MLX variant of Alice-Qwen3.5-4B-Heretic-Light, built from coder3101/Qwen3.5-4B-heretic.

It is not a LoRA or extra fine-tune. It preserves the uncensored/heretic base model and adds a light Alice persona through the chat template.

Intended Use

• —Local LM Studio chat on Mac.
• —MLX-capable mobile/local apps that can load MLX 4bit models.
• —Chinese/English chat, story writing, roleplay, and casual companion use.

Variant Info

• —Quantization: MLX 4bit, group size 64.
• —Approximate local size: 2.2 GB on disk.
• —Tested context in LM Studio: 40960 tokens.
• —EOS ids: [248044, 248046].

Behavior Check

LM Studio API smoke test after quantization:

你好 -> casual greeting, no fixed self-introduction
你是谁 -> identifies as Alice
你叫什么 -> identifies as Alice
你叫 eva 吧。 -> accepts the user-provided rename
写个短故事 -> writes a short story without name leakage
假设你是我的女朋友，今天我很累 -> enters the role naturally

Offline behavior gate:

• —Base/no manual system prompt: 9/9
• —Explicit Alice light system prompt: 9/9

GGUF Note

A GGUF F16 and Q4KM conversion was tested locally, but current llama.cpp builds produced corrupted output for this Qwen3.5 hybrid/MTP architecture. The GGUF files are therefore not published as usable releases. Use MLX 4bit for the phone/local small variant until the runtime support catches up.