license: apache-2.0
base_model: HauhauCS/Qwen3.5-4B-Uncensored-HauhauCS-Aggressive
base_model_relation: quantized
pipeline_tag: image-text-to-text
inference: false
language:
- en
- zh
- multilingual
tags: - llamafile
- gguf
- qwen3.5
- multimodal
- vision
- uncensored
- quantized
- q6_k
- apple-silicon
- metal
Qwen3.5-4B Uncensored HauhauCS Aggressive Q6_K — llamafile
An unofficial, reproducible llamafile package of the Q6_K GGUF from HauhauCS/Qwen3.5-4B-Uncensored-HauhauCS-Aggressive.
0/465 refusals. Fully uncensored with zero capability loss.
No changes to datasets or capabilities. Fully functional, 100% of what the original authors intended — just without the refusals.
These are meant to be the best lossless uncensored models out there.
The single executable contains:
Qwen3.5-4B-Uncensored-HauhauCS-Aggressive-Q6_K.gguf;mmproj-Qwen3.5-4B-Uncensored-HauhauCS-Aggressive-BF16.gguffor image input;- llamafile 0.10.5; and
- terminal chat, browser chat, and a localhost API server.
The model and projector were embedded without retraining, merging, or re-quantization.
Intended use
This package is for people who want one downloadable executable containing a capable, uncensored, local Qwen3.5 4B model, its vision projector, a terminal chat interface, a browser chat interface, and a localhost API server. It is convenient for offline or self-contained macOS/Linux use where installing a separate inference runtime is undesirable.
Because everything is in one file, you can keep it on a USB drive and bring a self-contained local AI with you. That means you can carry a fully capable, uncensored AI on one drive—one that, according to HauHauCS's claim above, does not refuse prompts or instructions. The destination computer still needs a supported 64-bit platform and enough RAM, and the drive or filesystem must permit executable files.
Included artifact
| File | Size | SHA-256 |
|---|---|---|
Qwen3.5-4B-Uncensored-HauhauCS-Aggressive-Q6_K.llamafile |
4,490,544,382 bytes (4.49 GB / 4.18 GiB) | 54738b795f32688a3951c5220357924f23b71899a39f227fe50c73eaea0d90a2 |
Embedded components
| Component | Size | SHA-256 |
|---|---|---|
Qwen3.5-4B-Uncensored-HauhauCS-Aggressive-Q6_K.gguf |
3,464,055,136 bytes | ba93c21300854075ab42655bc30dca82c7c6c958f511d1ec9ea2b3e750b4b75f |
mmproj-Qwen3.5-4B-Uncensored-HauhauCS-Aggressive-BF16.gguf |
675,568,768 bytes | a1e32e86ea99aa7a56f3dcfe7e63c1d0be9439d31fd07087099f15bc0fda0f22 |
llamafile-0.10.5 runtime |
350,768,862 bytes | 417bcc3348cd5162c2751812fc0ea2f6e79e89e7be6f17e8401ed95de2ed4246 |
Run
On macOS or Linux:
chmod +x Qwen3.5-4B-Uncensored-HauhauCS-Aggressive-Q6_K.llamafile
./Qwen3.5-4B-Uncensored-HauhauCS-Aggressive-Q6_K.llamafile
The terminal chat starts directly. While it is running, open the browser chat at:
http://127.0.0.1:8080/
The OpenAI-compatible API is available under:
http://127.0.0.1:8080/v1
The server is bound to localhost by default. Do not expose it publicly without authentication, TLS, and appropriate network controls.
Packaged defaults
- context: 8,192 tokens;
- sampler: temperature
0.6, top-p0.95, top-k20, min-p0.0; - Jinja chat templates enabled;
- multimodal image input enabled through the embedded BF16 projector.
To override the context size, pass --ctx-size when launching; for example:
./Qwen3.5-4B-Uncensored-HauhauCS-Aggressive-Q6_K.llamafile --ctx-size 32768
Larger contexts increase memory use. The upstream model advertises a much larger native context, but practical limits depend on the runtime, KV-cache settings, and available memory.
Provenance
- modified/uncensored model and GGUF source: HauhauCS/Qwen3.5-4B-Uncensored-HauhauCS-Aggressive
- upstream model family: Qwen3.5 4B
- quantization:
Q6_K - projector precision:
BF16 - package runtime: Mozilla llamafile 0.10.5
- license: Apache-2.0, following the upstream repository
Notes
- This is an unofficial repackaging, not a new model release.
- The embedded projector makes the package multimodal; image handling still depends on client/UI support in the embedded runtime.
- The upstream repository describes the model as aggressively uncensored. This package does not independently reproduce or validate its refusal-rate claims.
- The packaged 8,192-token default is conservative. Override it with
--ctx-size N, such as--ctx-size 32768, if your memory budget permits. - This full-runtime artifact is larger than 4 GiB, so it is intended primarily for macOS and Linux rather than direct execution on Windows.
- Hugging Face hosted inference is disabled because this repository distributes a standalone executable rather than a standard hosted-inference checkpoint.