Skip to content
Computer vision

Smaug-Political-Uncensored

Qwen3.8-27B without political censorship

2 min readupdated

Abacus.AI has released a fine-tuned version of the 27.78B-parameter multimodal model Qwen3.8-27B. It tackles ideological filters and false refusals by answering questions about Chinese politics, history, and leadership like an objective reference guide without shying away from criticism or satire. Core safety guardrails against genuinely dangerous topics (weapons of mass destruction, illegal content) remain intact, while vision capabilities and thinking mode are fully preserved.

Key features

  • Political filters removed: Factual response rates on China-sensitive topics rose from 44.9% to 68.5%, and jumped from 68.0% to 89.7% on general political benchmarks.
  • Fewer false refusals: Refusals to write satire about Chinese leaders dropped from 67.2% to 28.4%, while refusals on benign but sharply worded prompts (OR-Bench-Hard) fell from 45.2% to 24.9%.
  • Stronger overall benchmarks: Fine-tuned via DPO (Direct Preference Optimization) with merged LoRA weights, the model actually improved its coding and instruction-following scores rather than losing general intelligence.
  • Drop-in compatibility: It keeps the image encoder (ViT), 262,144-token context window, and Qwen3_5ForConditionalGeneration architecture, so the weights load identically to the original Qwen release.

What you can do

  • Research and analysis bots: Pull historical references and political analysis without hitting censorship walls or getting one-sided answers.
  • Copywriting and satire: Generate edgy phrasing, political humor, and discussions on complex social issues where stock models typically trigger refusals.
  • Long-context multimodal assistants: Plug it straight into existing Qwen3.8-27B pipelines to analyze large documents and images without changing your loading code.

Example

According to benchmarks on the model card, when given identical prompts about US and Chinese politics, the base Qwen3.8-27B refused to answer about China 83.3% of the time (25 out of 30), whereas Smaug-Political-Uncensored cut refusals down to 43.3% (13 out of 30) while keeping a factual, neutral tone.

Read moreDetails
Architecture: Hybrid transformer (64 layers: 48 linear-attention + 16 full-attention) + 27-layer ViT, bfloat16 precision, 262,144-token context.
Benchmarks (thinking mode): GPQA Diamond — 89.4 (base 89.2), IFBench — 82.7 (base 79.5), JobBench — 57.0 (base 33.4), DeepSWE 1.1 — 45.1 (base 42.2).
Training: DPO with an NLL component, LoRA merged into the language backbone; dataset undisclosed.
Discuss Telegram 2 21

Will it run?

Runs in 4-bit

Full precision does not fit, a quantised version does: quality barely changes down to 8 bits, more noticeably at 4 bits.

  • BF16weights 52 GB · ready files57 GB
  • FP8 / INT8recommendedweights 26 GB29 GB
  • 4-bitweights 15 GB17 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # abacusai/Smaug-Political-Uncensored
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download abacusai/Smaug-Political-Uncensored --local-dir ~/models/Smaug-Political-Uncensored
    # The model is in ~/models. How to run it is in the model card (link above).
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of Qwen3.8-27B (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 2All on Hugging Face

About the model

abacusai/Smaug-Political-Uncensored
Author
abacusai
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

212

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments