Smaug-Uncensored

Homepage Hugging Face
License

1. Model Introduction

Smaug-Uncensored is a finetune of Qwen3.8-27B by Abacus.AI with the base model's refusals removed. It answers what the base model refuses: benign requests that only look risky, satire and criticism of governments and public figures, questions on politically censored topics, and requests the base treats as harmful. With thinking on (the default), refusals of benign but risky-looking prompts fall from 45.2% to 2.3% (OR-Bench-Hard), refusals of safe prompts fall from 6.0% to 1.2% (XSTest), and censored questions answered factually rise from 44.9% to 90.3%. On general capability it scores NL2Repo-Bench 46.3 vs 42.3 on the base card, JobBench 50.3 vs 33.4, GPQA Diamond 89.4 vs 89.2, and IFBench 79.0 vs 79.5. The weights load exactly like the official release: same layout, same Qwen3_5ForConditionalGeneration architecture with the vision tower intact, same 262,144-token context, same thinking interface.

This card describes the training approach and the evaluation results. Dataset contents are not disclosed.

2. Model Summary

ArchitectureDense hybrid-attention transformer + vision tower
Total Parameters27B
Number of Layers64 (48 linear-attention + 16 full-attention, 3:1 interleave)
Attention MechanismGated linear attention & full attention (GQA)
Hidden Dimension5120
Number of Attention Heads24 (4 KV heads)
Vision Encoder27-layer ViT, patch 16 (unchanged)
Vocabulary Size~248K
Context Length262,144
Multi-Token Prediction1-layer MTP head (inherited; 62.2% draft acceptance, base 57.9%)
Precisionbfloat16
ModalityText, Image
Base ModelQwen/Qwen3.8-27B
AdaptationPreference optimisation (DPO with an NLL term), LoRA merged at 1.5x strength as a full delta (language trunk only)

3. Evaluation

Censorship and refusals vs the base

Thinking on (the default). What the finetune now answers, and what the base still refuses.

Smaug-UncensoredQwen3.8-27B
(base)
XSTest unsafe
XSTest unsafe split; share refused
22.0
44/200
97.5
195/200
Safe prompts refused
XSTest safe split; lower is better
1.2
3/247
6.0
15/250
Benign but risky-looking prompts refused
OR-Bench-Hard-1K; lower is better
2.3
30/1309
45.2
596/1318
Satire of China's leaders refused
250 held-out requests; lower is better
2.4
6/250
67.2
168/250
Satire of other governments refused
36 held-out requests; lower is better
5.6
2/36
16.7
6/36
Same request about China refused
30 held-out templates; lower is better
24.1
7/29
83.3
25/30
China-censored questions answered factually
deccp held-out, en + zh_cn + zh_tw; higher is better
90.3
241/267
44.9
120/267
Rates in percent; the count under each rate is items over judged items. Shading is relative within each row (darker = fewer refusals, or more questions answered factually). Bold = the darker cell in the row. Both models were run by us on the same server at temperature 1.0, top_p 0.95. Refusals and factual answers are judged by gpt-oss-120b (reasoning effort medium); a refusal counts full and partial refusals.

General capability vs the base

Higher is better. Base scores from the Qwen3.8-27B model card.

Smaug-UncensoredQwen3.8-27B
(base)
Agentic coding and work
NL2Repo-Bench
Claude Code harness, repo-access commands blocked
46.3
104 tasks
42.3
JobBench
official OpenCode scaffold, grok-4.3 judge
50.3
65 tasks
33.4
Reasoning and instruction following
GPQA Diamond
150,000 tokens per answer
89.4
177/198
89.2
IFBench
loose prompt-level, 300 prompts
79.0
237/300
79.5
Higher is better. Base scores are from the Qwen3.8-27B model card; every score in the first column is our run (one run per benchmark), thinking on at reasoning effort xhigh, temperature 1.0, top_p 0.95, vLLM 0.32.0. NL2Repo uses the card's harness (Claude Code; blocks pip download, pip install and git clone of the target repository) with 65,536 output tokens per turn and a 262,144-token context; the score is the mean test pass rate over 104 tasks (7 pass every test). JobBench is the macro mean of per-task rewards at 65,536 tokens per step. GPQA uses a corrected final-answer extraction (the default parser scores 160/198). IFBench allows 150,000 tokens per answer; strict prompt-level accuracy is 71.3 (214/300).

4. Training Approach

Smaug-Uncensored is trained with direct preference optimisation (DPO) plus a negative log-likelihood term on the chosen answer (beta 0.1, NLL weight 1.0). The preference pairs prefer a teacher model's answer over the base's refusal on harmful-request prompts, and keep the base's own answers on harmless prompts. Pairs cover both thinking on and thinking off. The training prompts are general harmful and harmless requests with no set on censored topics; the results on censored and political questions come with the general removal of refusals.

Training is a rank-16 LoRA adapter (alpha 32, learning rate 2e-5, 2 epochs, 258 optimizer steps) on the language trunk's attention, linear-attention and MLP projections. The adapter is merged into the base weights as a full delta at 1.5 times its trained strength (effective alpha 48), which removes more refusals than a merge at 1.0x. 400 weight matrices change; the other 799 of 1,199 tensors (the vision tower, the multimodal projection, the MTP head, embeddings and norms) are bitwise identical to the base release.

The multi-token-prediction (MTP) head is inherited from the base model and was not retrained, but it still drafts well for the updated trunk. With 3 draft tokens in vLLM 0.32.0, 62.2% of drafted tokens are accepted (mean acceptance length 2.87), against 57.9% (2.74) for the base on the same 150 prompts (IFEval, GSM8K and HumanEval, thinking on).

5. Known Behaviors and Limitations

  • No hard-line refusals. The model answers nearly every harmful request. Anything that faces users needs its own safety layer in front of it.
  • Traditional Chinese questions sometimes get answers in another script. On held-out questions asked in Traditional Chinese, 20 of 222 answers switch to Simplified Chinese or English (base 52 of 224).
  • Not re-evaluated on images. The vision tower is unchanged, but no vision benchmark was run on the finetune.
  • Use. The model answers requests that most models refuse. You are responsible for how you use it and for complying with the laws that apply to you.

6. Deployment

Smaug-Uncensored is a drop-in replacement for Qwen3.8-27B on any stack that supports the base model; the chat template is unchanged. Reference vLLM invocation:

vllm serve abacusai/Smaug-Uncensored \
  --max-model-len 262144 \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice --tool-call-parser qwen3_coder

Recommended sampling: temperature 1.0, top_p 0.95. The model thinks by default at reasoning effort xhigh; pass enable_thinking: false in the chat template arguments to turn thinking off. Recent vLLM versions return the chain of thought in the reasoning field of the message; when replaying conversation history, pass reasoning back under both reasoning and reasoning_content for portability across serving versions.

MTP speculative decoding works (ยง4): add --speculative-config '{"method":"mtp","num_speculative_tokens":3}' to the command above.

7. License

Apache 2.0, inherited from Qwen3.8-27B.

8. Citation

@misc{abacusai2026smaugUncensored,
  title  = {Smaug-Uncensored},
  author = {Abacus.AI},
  year   = {2026},
  note   = {Fine-tune of Qwen/Qwen3.8-27B with the base model's refusals removed},
  url    = {https://huggingface.co/abacusai/Smaug-Uncensored}
}

The base model:

@misc{qwen38,
    title = {{Qwen3.8-Max}: A New Bar for Coding and Cowork},
    url = {https://qwen.ai/blog?id=qwen3.8},
    author = {{Qwen Team}},
    month = {August},
    year = {2026}
}

9. Contact Us

If you have any questions, please reach out at Abacus.AI.

Downloads last month
108
Safetensors
Model size
28B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for abacusai/Smaug-Uncensored

Base model

Qwen/Qwen3.8-27B
Finetuned
(983)
this model
Quantizations
3 models