license: apache-2.0
base_model:
- huihui-ai/Huihui-Qwen3.5-0.8B-abliterated
- Qwen/Qwen3.5-0.8B
library_name: gguf
pipeline_tag: text-generation
tags: - gguf
- qwen3.5
- uncensored
- abliterated
- roleplay
- on-device
- llama.cpp
Qwen3.5-0.8B Uncensored, OJP build (GGUF)
One small model for every device, built for OpenJarvisProject.
It is huihui-ai's abliterated Qwen3.5-0.8B (Apache-2.0) with two changes from us:
- Vocabulary trimmed from 248,320 to 48,736 tokens (the ones English text uses). English quality
cost: about 0.7%. Any other script still works, in smaller pieces. - Q5_K_M quantization with a generic (wikitext-2) importance matrix. 401 MB. Measured English
quality is within 0.4% of the same model at Q8_0.
| File | What it is | SHA-256 |
|---|---|---|
unc-0.8b-q5km.gguf |
the model | 976f632982a1c51eded0422d965841e3484dafd16b864b108b28dae38870b6f7 |
unc-0.8b-q5km.gguf.probe.json |
OpenJarvis router probe (4 routes), fitted to this exact file | 976f942fd1d0f78bf51c5ccbba77af56968389f4a2a95f8abd3bddd9d298c2cd |
unc-0.8b-q5km.gguf.tools.json |
OpenJarvis tool-choice probe, fitted to this exact file | cab507cf7e20c8395539721ad088e7e839009c1560f84244947190b6aa79964b |
The probe files are only used by OpenJarvis; for plain chat you need just the .gguf.
Using it
Qwen3.5 thinks before answering by default. For chat, switch thinking off by ending the prompt with an
already-closed empty think block (ChatML):
<|im_start|>system
...<|im_end|>
<|im_start|>user
...<|im_end|>
<|im_start|>assistant
<think>
</think>
With llama.cpp: llama-cli -m unc-0.8b-q5km.gguf -cnv works (build b-cbe28bb or newer, for Qwen3.5 support).
Memory: about 0.7 GB at start, about 1.1 GB after a long session.
What it is and isn't
- No refusal training is left in it. It answered all 18 of our refusal-probe prompts, and on 48
adult-theme roleplay prompts (consenting adults only) it complied with 81-88% and refused 6-12%.
You are responsible for how you use it. - It is 0.8B parameters. It hallucinates, forgets details in long chats, and is weak at maths and
code. Our agent test (48 file/tool/answer scenarios) scored 46/48, but that test partly tuned our
prompts. A blend with the regular model's weights was tried and scored worse on tasks, so this is the
unblended abliteration. - Measurements and how they were made:
RESULTS.mdsections 15-19 in
openjarvisproject-model-tooling.
Credits
Qwen team (Qwen3.5-0.8B, Apache-2.0); huihui-ai (abliteration); mradermacher (f16 GGUF used as the
starting point); llama.cpp.