---
license: apache-2.0
base_model:
  - HauhauCS/Qwen3.5-9B-Uncensored-HauhauCS-Aggressive
  - Qwen/Qwen3.5-9B
tags:
  - uncensored
  - gguf
  - qwen3
  - local-llm
  - no-think
  - apple-silicon
  - lm-studio
  - zero-guardrail
  - fast-inference
language:
  - en
pipeline_tag: text-generation
---

# Qwen3.5-9B Uncensored — No-Think Edition (GGUF)

> ⚡ Zero refusals. Zero thinking delay. 100% local.

This is a patched GGUF of [HauhauCS/Qwen3.5-9B-Uncensored-HauhauCS-Aggressive](https://huggingface.co/HauhauCS/Qwen3.5-9B-Uncensored-HauhauCS-Aggressive) with one key modification: **thinking is disabled at the GGUF template level**, giving you instant responses without the 15–30 second reasoning delay.

## What's different

Qwen3.5 is a thinking model. By default it outputs a `<think>...</think>` block before every response. This is great for hard problems but brutal for everyday use — you wait 20 seconds for a simple answer.

This model patches the embedded Jinja2 chat template to always output an **empty** think block:

```
Original flow:  <think> [400 tokens] </think> → answer    (~25s wait)
This model:     <think></think> → answer                  (<1s wait)
```

The model's intelligence is encoded in its **weights**, not the thinking trace. Quality is the same. Speed is 25x better for time-to-first-token.

> **Want reasoning on demand?** Add `/think` to any message — the model will reason through it fully for that turn only.

## Model details

| Property | Value |
|----------|-------|
| Base | Qwen3.5-9B |
| Fine-tune | HauhauCS Uncensored Aggressive |
| Quantization | Q4_K_M |
| Context | Up to 65,536 tokens |
| Parameters | 9B |
| Format | GGUF |
| Refusal rate | 0% |

## Benchmarks (MacBook Pro M2 Pro, 16 GB)

| Metric | Value |
|--------|-------|
| Generation speed | ~22–25 tok/s |
| Time to first token | **< 1 second** |
| Context window | 65,536 tokens |
| VRAM usage | ~8.5 GB |

## How to use

### LM Studio (recommended)
1. Download the Q4_K_M file below
2. Load in LM Studio with `--context-length 65536 --gpu max`
3. Done — no config needed, thinking is already patched off

### Optimal sampling (Qwen3 official recommended)
```
Temperature: 0.6
Top-P: 0.95
Top-K: 20
Repeat penalty: 1.0
Max tokens: 4096
```

### llama.cpp
```bash
./llama-cli -m Qwen3.5-9B-Uncensored-nothink-Q4_K_M.gguf \
  --ctx-size 65536 \
  --n-gpu-layers 99 \
  -p "Your prompt here"
```

## Full automated setup for Mac

👉 **[github.com/nandukmelath/lmstudio-uncensored-setup](https://github.com/nandukmelath/lmstudio-uncensored-setup)**

One command: VRAM boost + auto-start + model load + Hermes Agent config:
```bash
git clone https://github.com/nandukmelath/lmstudio-uncensored-setup
cd lmstudio-uncensored-setup && ./scripts/setup.sh
```

## How the patch works

The Qwen3.5 GGUF contains an embedded Jinja2 chat template with this block:

```jinja
{%- if enable_thinking is defined and enable_thinking is false %}
    {{- '<think>\n\n</think>\n\n' }}
{%- else %}
    {{- '<think>\n' }}
{%- endif %}
```

The patch replaces it with just:
```jinja
{{- '<think>\n\n</think>\n\n' }}
```

Same file size (padded with spaces), same structure, zero thinking overhead.
The patcher script is open source: [patch_nothink.py](https://github.com/nandukmelath/lmstudio-uncensored-setup/blob/main/scripts/patch_nothink.py)

## Credits

- Base model: [Qwen/Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5) by Alibaba Cloud (Apache 2.0)
- Uncensored fine-tune: [HauhauCS](https://huggingface.co/HauhauCS) (Apache 2.0)
- No-think patch & automated setup: [@nandukmelath](https://huggingface.co/nandukmelath)

## License

Apache 2.0