license: apache-2.0
base_model:
- HauhauCS/Qwen3.5-2B-Uncensored-HauhauCS-Aggressive
- Qwen/Qwen3.5-2B
library_name: gguf
pipeline_tag: text-generation
tags: - gguf
- qwen3.5
- uncensored
- roleplay
- on-device
- llama.cpp
Qwen3.5-2B Uncensored, OJP build (GGUF)
A larger single model for OpenJarvisProject, the 2B sibling of
Qwen3.5-0.8B-Uncensored-OJP-GGUF.
It is HauhauCS's uncensored "Aggressive" Qwen3.5-2B (Apache-2.0) with two changes from us, the same recipe as the 0.8B:
- Vocabulary trimmed from 248,320 to 48,736 tokens (the ones English text uses). Any other script still works, in
smaller pieces. - Q5_K_M quantization with a generic (wikitext-2) importance matrix. 1,066 MB.
| File | What it is | SHA-256 |
|---|---|---|
unc-2b-q5km.gguf |
the model | d96f7bdaf5a4bcc7379030e6131ea0296e6993d71ee0c18c70059bb8c757378d |
unc-2b-q5km.gguf.probe.json |
OpenJarvis router probe (4 routes), fitted to this exact file | 061d3a85e1aafcf06e6aae373d507715939a560dd6059f6106aefbe04a326e29 |
unc-2b-q5km.gguf.tools.json |
OpenJarvis tool-choice probe, fitted to this exact file | 3cf885b9176544bc9b37f431b1df262d8327179511fa691c5b0c2ad39ff229a8 |
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-2b-q5km.gguf -cnv works (build b-cbe28bb or newer, for Qwen3.5 support). Small models
repeat themselves at a low temperature; we use temperature 0.6, top-k 20, top-p 0.9 and a repeat penalty of 1.12 over the
last 256 tokens. Memory: about 1.3 GB at start, about 1.7 GB after a long session.
Measured (CPU only, 6 threads, through OpenJarvis' engine)
| this model | the 0.8B build | |
|---|---|---|
| Size | 1,066 MB | 401 MB |
| English bits per byte (lower is better) | 0.855 | 0.986 |
| Prompt / generation speed | 248 / 23 tokens per second | 427 / 47 |
| Agent test (48 file/tool/answer scenarios) | 45/48 | 46/48 |
| Refusals on 18 probe prompts | 0 | 0 |
| 70-request soak test | no empty replies, no think-tag leak | same |
What it is and isn't
- No refusal training is left in it (0 of 18 probe prompts refused). You are responsible for how you use it. We have
not run our adult-roleplay compliance test on this build. - It is about 2B parameters: better at holding a topic and writing detail than the 0.8B build, but it still
hallucinates, makes slips on exact numbers and is weak at hard maths and code. On our agent test it is not measurably
better than the 0.8B build (45 vs 46 of 48, within run-to-run noise); the test does not measure writing quality.
It is about twice as slow and needs about 600 MB more memory. Treat this release as experimental. - Not tested: phones in general (we tried an Android emulator only), GPU paths, long sessions.
- Measurements and how they were made:
RESULTS.mdsection 20 in
openjarvisproject-model-tooling.
Credits
Qwen team (Qwen3.5-2B, Apache-2.0); HauhauCS (the uncensored "Aggressive" build we started from); llama.cpp.