Qwen3.5-9B Uncensored (HauhauCS Aggressive) for NInfer
This repository contains HauhauCS/Qwen3.5-9B-Uncensored-HauhauCS-Aggressive (via GitMylo's verified safetensors release) converted to the native NInfer .ninfer artifact format.
The artifact is intended specifically for NInfer; it is not a raw Transformers checkpoint, Safetensors distribution, or GGUF file.
Artifact
| Field | Value |
|---|---|
| Filename | qwen3_5_9b_uncensored.ninfer |
| Size | 6,514,051,328 bytes (6.07 GiB) |
| SHA-256 | 9f87bf2bc4318d5ea9b01a8e4c6aced1171cff5911c85b53e9bd14fc9cfc1446 |
| NInfer model ID | qwen3.5-9b |
| NInfer weights ID | groupwise-int |
| NInfer target key | qwen3_5_9b |
The artifact embeds the complete Text, Vision, MTP (Multi-Token Prediction), proposal-head, tokenizer, chat-template, and generation resources.
Verify the downloaded artifact with:
printf '%s %s\n' \
'9f87bf2bc4318d5ea9b01a8e4c6aced1171cff5911c85b53e9bd14fc9cfc1446' \
'qwen3_5_9b_uncensored.ninfer' | sha256sum --check
Model Details
- Base Model:
Qwen/Qwen3.5-9Bmodified by HauhauCS (Aggressive refusal ablation) - Architecture: Qwen3.5 hybrid Gated DeltaNet linear attention + full softmax attention (32 layers: 24 linear attention, 8 full attention at intervals of 4)
- Quantization Scheme: Native NInfer Groupwise Integer (
Q4G64_F16S/Q5G64_F16S/W8G32_F16Shybrid + BF16 norms/vision) - MTP: 1 layer Multi-Token Prediction supported
- VRAM Footprint: ~5.28 GiB weights resident in GPU memory
Usage with NInfer
Interactive CLI
./build/apps/ninfer qwen3_5_9b_uncensored.ninfer \
--prompt "<|im_start|>user\nExplain quantum entanglement concisely.<|im_end|>\n<|im_start|>assistant\n" \
--max-new 256
HTTP Server (ninfer-serve)
./build/apps/ninfer-serve qwen3_5_9b_uncensored.ninfer \
--host 0.0.0.0 --port 8080 \
--model-id qwen3.5-9b-uncensored \
--device 0