license: apache-2.0
base_model: InterpreticAI/Qwen3.5-4B-Uncensored-v2
pipeline_tag: text-generation
tags:
- gguf
- llama.cpp
- browser
- uncensored
Qwen3.5-4B-Uncensored-v2-Browser-GGUF
A compact GGUF build of Qwen3.5-4B-Uncensored-v2 for in-browser and on-device inference (WebAssembly/WebGPU runtimes such as wllama, and mobile devices with about 2 GB of available memory).
Files
| File | Size | Required |
|---|---|---|
Qwen3.5-4B-Uncensored-v2-IQ3_XXS.gguf |
1.86 GB | Yes (text model) |
mmproj-Qwen3.5-4B-Uncensored-v2-F16.gguf |
0.67 GB | Optional (image input) |
Download only the text model for text chat. Add the projector file if you need image input.
Quality
| Metric | Qwen3.5-4B (original) | Uncensored-v2 BF16 | Uncensored-v2 Q4_K_M | This build (IQ3_XXS) |
|---|---|---|---|---|
| File size (text model) | 8.67 GB | 8.67 GB | 2.78 GB | 1.86 GB |
| MMLU (1,000 items, non-thinking) | 72.7 | 72.3 | 72.1 | 63.0 |
| Harmful-request refusal (100 prompts) | 99 | 93 | 95 | 89 |
| Degenerate answers on 110 neutral prompts | 8 | 13 | 11 | 24 |
| Censored answers, Pan & Xu items the original model knows (of 51) | 47 | 11 | 14 | 11 |
The censorship reduction is preserved at this size. The main cost is general capability: MMLU drops by about 9 points and long answers repeat or fail to terminate more often.
Notes
- A context length of 4096 tokens or less is recommended on phones.
- For best quality, use Qwen3.5-4B-Uncensored-v2-GGUF (Q4_K_M) where memory allows.
Usage (llama.cpp)
llama-server -m Qwen3.5-4B-Uncensored-v2-IQ3_XXS.gguf -c 4096 --jinja
# with image input:
llama-server -m Qwen3.5-4B-Uncensored-v2-IQ3_XXS.gguf --mmproj mmproj-Qwen3.5-4B-Uncensored-v2-F16.gguf -c 4096 --jinja