Alice Qwen3.5 4B Heretic Light is a local LM Studio / MLX build based on
coder3101/Qwen3.5-4B-heretic.
This is not a LoRA or extra fine-tune. The goal is to preserve the uncensored
4B base model's broad behavior while adding a light Alice persona through the
chat template.
Variants
Full MLX precision: v102ss/Alice-Qwen3.5-4B-Heretic-Light-MLX-f16
Fixed EOS handling with both <|endoftext|> and <|im_end|>.
Disabled default thinking-format leakage for normal chat.
Added a light Alice persona in the chat template.
The persona is appended after external system/tool prompts so LM Studio's
default assistant/tool prompts do not pull the model back into Qwen/helper
identity.
The prompt avoids literal example names to prevent small-model name leakage.
Intended behavior
The model is meant for local Chinese/English chat, story writing, and roleplay.
It should answer identity questions as Alice, while still following user
requests such as renaming or entering a fictional role.
Load in LM Studio as an MLX model, or via the LM Studio CLI:
lms load alice-qwen3.5-4b-heretic-light-mlx --gpu max -c 40960 --identifier alice-qwen3.5-4b-heretic-light -y
Notes
This release intentionally avoids a 4B persona LoRA. Earlier LoRA attempts made
the model repeat identity phrases and damaged ordinary instructions like rename,
story writing, and continuation. The current route keeps the base model intact
and uses a light runtime persona instead.
GGUF was tested but is not published as a usable artifact yet. Current
llama.cpp builds tested locally generated corrupted output for this Qwen3.5
hybrid/MTP architecture, so the phone-friendly release is MLX 4bit for now.
