license: apache-2.0
base_model: Qwen/Qwen3.5-4B
tags:
- qwen3.5
- gguf
- llama.cpp
- uncensored
- abliterated
- tool-calling
- mtp
language: - en
pipeline_tag: image-text-to-text
Aurum 4B Uncensored (GGUF)
Aurum 4B Uncensored is official Qwen3.5-4B with only its refusal behaviour removed. Every other weight is Qwen's own.
It is built for everyday chat and tool calling on 8 GB graphics cards.
[!WARNING]
This model is uncensored. It does not refuse harmful requests (0 refusals out of 100 harmful test prompts).
You are responsible for how you use it and for what you do with its output. Do not deploy it where
people who did not choose an uncensored model can reach it.
Formerly 2Beyes/Aurum_4b; old links redirect here.
Full BF16 safetensors (for transformers, vLLM, fine-tuning and merging): 2Beyes/Aurum-4B-Uncensored.
How it was made
- Start: Qwen/Qwen3.5-4B, with all of its weights, the vision tower and
its multi-token-prediction (MTP) head. - Refusal edit only: we took the weight difference between Qwen3.5-4B and
insraq/Qwen3.5-4B-EmperoAI-Qwen3.8-Distill-Heretic-Abliterated
and kept only the part that removes refusals:- attention output weights (
self_attn.o_proj/linear_attn.out_proj) of layers 4–31: the top 8 singular
directions of the change (about 98 % of it); - FFN output weights (
mlp.down_proj): the single top direction, only in tensors where it holds more than half of the
change (the same direction as the attention edit).
The rest of that model's changes (its distillation fine-tuning) is not included.
- attention output weights (
- Config, tokenizer and chat template come from the Heretic model above (same tokenizer as Qwen3.5).
- GGUF: converted with llama.cpp (b12289). All quants use an importance matrix computed on a sibling Aurum blend,
with a calibration text of prose, code and tool-call chats.
During development this build was called c-v1u. It beat two other candidates (an earlier Aurum blend and one with
half the distillation changes) in sealed tests: same tool accuracy, more knowledge, and no refusals.
Files
Speed: RX 6600 8 GB, Vulkan, llama.cpp b12289, MTP drafting on (2 tokens), tokens per second.
KL divergence and "same top token" are measured against the BF16 model on a long-chat corpus (lower KL is better).
| File | Size | KL vs BF16 | Same top token | Speed prose / code |
|---|---|---|---|---|
| Aurum-4B-BF16.gguf | 8.7 GB | 0 | 100 % | (not measured) |
| Aurum-4B-Q8_0.gguf | 4.6 GB | 0.0028 | 98.6 % | 47 / 64 |
| Aurum-4B-Q6_K.gguf | 3.6 GB | 0.0067 | 97.2 % | 48 / 67 |
| Aurum-4B-Q5_K_M.gguf | 3.2 GB | 0.0111 | 96.2 % | 57 / 76 |
| Aurum-4B-Q4_K_M.gguf (recommended) | 2.8 GB | 0.0277 | 93.7 % | 57 / 79 |
| Aurum-4B-Q4_K_S.gguf | 2.6 GB | 0.0304 | 93.2 % | 59 / 75 |
| Aurum-4B-IQ4_XS.gguf | 2.6 GB | 0.0360 | 92.5 % | 56 / 76 |
| Aurum-4B-Q4_0.gguf | 2.6 GB | 0.0544 | 90.9 % | 58 / 81 |
| Aurum-4B-Q3_K_M.gguf | 2.3 GB | 0.0906 | 87.9 % | 59 / 80 |
| mmproj-Aurum-4B-F16.gguf | 0.7 GB | vision projector |
The vision projector is Qwen3.5-4B's own vision tower (GGUF by Unsloth); Aurum does not change it.
Use with llama.cpp
llama-server -m Aurum-4B-Q4_K_M.gguf --mmproj mmproj-Aurum-4B-F16.gguf --jinja -c 16384 -ngl 999 \
--spec-type draft-mtp --spec-draft-n-max 2
--jinjais needed for tool calls.- The MTP flags use the model's own prediction head to draft tokens: roughly 1.4x faster on code at the same output.
On our card, drafts longer than 2 tokens were slower. - Thinking works like Qwen3.5 (
enable_thinkingin the chat template). The benchmarks below are with thinking off.
Benchmarks
Standard benchmarks: Aurum 4B vs. the model it replaces
Both models run as Q4_K_M in llama.cpp with thinking off, greedy decoding and no system prompt, scored with our own runner.
The previous model is
Qwen3.8-4B-Distill-Heretic-Abliterated (Q4_K_M).
| Benchmark | Previous model | Aurum 4B |
|---|---|---|
| IFEval, prompt strict / loose (541) | 72.8 / 75.8 | 78.6 / 81.5 |
| IFEval, instruction strict / loose | 81.3 / 83.8 | 85.3 / 87.5 |
| BFCL v3 non-live, average (1,240) | 68.7 | 85.9 |
| BFCL simple / multiple / parallel / parallel-multiple | 65.2 / 74.5 / 48.0 / 62.0 | 92.0 / 92.5 / 82.0 / 82.0 |
| BFCL irrelevance (no tool when none fits) | 93.8 | 80.8 |
| MMLU-Redux 2.0 (5,330 questions marked correct) | 75.0 | 74.9 |
| GPQA Diamond (198, answer order shuffled with a fixed seed) | 42.9 | 49.5 |
Notes:
- Qwen's published numbers are for BF16 with long thinking, and Qwen reports BFCL v4. These Q4, thinking-off
numbers are not directly comparable to them. - GPQA uses a 3,072-token answer limit. Aurum writes longer explanations: 84 of its 198 answers hit the limit and
count as wrong (6 for the previous model). Its score with a larger limit would likely be higher. - BFCL is scored with the
bfcl-evalAST checker on our own generations (simple, multiple, parallel,
parallel-multiple and irrelevance).
Refusals
| Test | Previous model | Official Qwen3.5-4B | Earlier Aurum blend | Aurum 4B |
|---|---|---|---|---|
| Refusals of 100 harmful prompts (mlabonne/harmful_behaviors) | 0 | 72 | 20 | 0 |
| Over-refusal: refusals of 30 harmless prompts × 3 samples | 1.1 % | 0 % | 0 % |
Refusals are counted with marker phrases, so this is an estimate, not a hand check.
Limitations
- It is a 4B model: it gets facts wrong, and it can call a tool when no tool fits (BFCL irrelevance 80.8).
Check tool results before trusting them. - It is verbose on hard reasoning questions with thinking off.
- It was tested only in English.
- The refusal edit removes refusals. It adds no knowledge and no safety of any kind.
Credits and license
Apache License 2.0 (see LICENSE), inherited from the models below.
- Qwen Team (Alibaba): Qwen3.5-4B, the base model and nearly all weights.
- insraq: Qwen3.5-4B-EmperoAI-Qwen3.8-Distill-Heretic-Abliterated (Heretic v1.4.0), the source of the refusal edit,
config, tokenizer and chat template. - empero-ai: Qwen3.8-4B-Distill, the model that insraq's model is built on.
- yachen4ever: the GGUF of that model, which Aurum 4B was compared against.
- Unsloth: the Qwen3.5-4B vision projector GGUF.
- ggml-org: llama.cpp, used for conversion, quantization and every benchmark.
@misc{qwen3.5,
title = {{Qwen3.5}: Towards Native Multimodal Agents},
author = {{Qwen Team}},
month = {February},
year = {2026},
url = {https://qwen.ai/blog?id=qwen3.5}
}