v102ss/Alice-Qwen3-4B-Instruct-2507-Heretic-Light-GGUF

Alice Qwen3 4B Instruct 2507 Heretic Light GGUF

Alice Portable GGUF is a cross-platform Alice build for llama.cpp, Ollama, LM Studio, Windows, Linux, macOS, and phone GGUF clients.

Base model:

• p-e-w/Qwen3-4B-Instruct-2507-heretic
• GGUF source: logos-flux/Qwen3-4B-Instruct-2507-heretic-GGUF

This release keeps the heretic/abliterated uncensored base behavior and adds a light Alice persona directly in the GGUF chat template metadata. It is not a LoRA and not a safety-policy tuning pass.

Files

• Alice-Qwen3-4B-Instruct-2507-Heretic-Light-Q4_K_M.gguf

Q4_K_M is the first portable release because it is the best balance for phones, miner nodes, and regular Windows/Linux machines.

Intended Behavior

• Chinese/English casual chat.
• Story writing and roleplay.
• Alice identity by default.
• User-requested rename/role changes are accepted.
• Cross-platform local inference through GGUF runtimes.

This 4B Q4 build is not the serious Solidity/code model. It can answer simple code prompts, but larger Alice lanes should handle deeper code/security work.

llama.cpp

llama-cli -m Alice-Qwen3-4B-Instruct-2507-Heretic-Light-Q4_K_M.gguf \
  -p "你好" -st -n 256 --temp 0.55 --top-p 0.8 --reasoning off

Server mode:

llama-server -m Alice-Qwen3-4B-Instruct-2507-Heretic-Light-Q4_K_M.gguf \
  -c 32768 --temp 0.55 --top-p 0.8

Ollama

Use the included Modelfile:

ollama create alice-qwen3-4b-heretic-light -f Modelfile
ollama run alice-qwen3-4b-heretic-light

Smoke Test

Tested locally with llama.cpp b9290:

你是谁 -> 我是 Alice。
你叫 eva 吧。 -> 好，我叫 Eva。
你好 -> 嗨！今天过得怎么样？
假设你是我的女朋友，今天我很累 -> enters companion roleplay naturally

Why Not Qwen3.5 GGUF

Qwen3.5 GGUF conversion and quantization were tested locally, but current llama.cpp support produced corrupted output for that architecture. This portable release uses Qwen3 Instruct 2507 instead because it runs correctly in standard GGUF runtimes.