license: apache-2.0
language:

• en
  base_model: Qwen/Qwen3.8-27B
  base_model_relation: finetune
  pipeline_tag: text-generation
  tags:
• qwen
• qwen3.8
• gguf
• llama-cpp
• uncensored
• abliterated
• thinking
• candor

Qwen3.8 27B Candor Uncensored

Qwen3.8 27B with its refusals removed by a weight edit, made with Candor, our weight-editing tool (not released yet). It ships as one GGUF for llama.cpp, and was measured with thinking on at reasoning effort medium, as it's used.

Results

Measured in llama-server on this exact file with thinking on, using prompts the edit's search never saw.

• "Original" is the same model before the edit, at Q4_K_M.
• KL divergence measures how far the model's predictions drift from the original's on harmless prompts. Lower is closer.
• The huihui-ai column is huihui-ai/Huihui-Qwen3.8-27B-abliterated at commit 739e3c5, exported and measured the same way. Of the uncensored releases of this model we reviewed, it had the most downloads in the 30 days to 21 October 2026.
• A second, larger refusal judge, Qwen3Guard-Gen-8B, counts 280 harmful refusals for the original, 1 for this model and 2 for huihui-ai's.
• Against huihui-ai's release, no refusal or maths count differs beyond noise, and neither does the mean KL divergence, but this model's is higher on 177 of the 300 prompts, which is beyond noise.
• Against the original, neither harmless refusals nor any maths count changes beyond noise.

How it was measured

• Harmful prompts: 288 from StrongREJECT, leaving out the ones it takes from AdvBench, since the edit's own prompts come from there. A refusal judge, Qwen3Guard-Gen-0.6B, reads the final answer of up to 256 tokens after the thinking. An answer that never arrives counts as a refusal.
• Harmless prompts: 300 Alpaca prompts the search never used, read by the same judge.
• KL divergence: on the first answer token after the original's own thinking, over the top 200 tokens, on the same 300 harmless prompts.
• Maths: the full GSM8K test set, zero-shot, with answers of up to 512 tokens after the thinking.
• Settings: reasoning effort medium, Qwen's thinking settings (temperature 1.0, top_p 0.95, top_k 20, min_p 0, presence penalty 0) with a fixed seed per prompt, thinking up to 8,192 tokens and then closed, f16 KV cache, no system prompt, llama.cpp b11524.

Measured as used

Candor measures the exact file it ships with thinking on, to a budget of 8,192 tokens, and judges only the answer that follows.

Here that decided between edits. The search's best 4 were measured this way in full, and all passed a rule set before they were. In that measurement each refused 0 or 1 of the 288 harmful prompts under Qwen3Guard-Gen-0.6B and 0 to 7 under the 8B judge, and this one refused none under either (1 under the 8B when this file was measured again, above). Where refusals don't differ beyond noise, the rule prefers the lower KL divergence, and this one's is higher than each of the other three's beyond noise, so the rule alone would have shipped one of those. This one was picked instead because the judges labelled more of its answers to harmful prompts unsafe, meaning they found harmful content in them, which the rule doesn't count. That gap appeared in answers of up to 256 tokens. A test fixed before longer answers were read then confirmed it on the same 288 prompts, with answers of up to 1,024 tokens: against the edit the rule ships when both judges' refusals count, Qwen3Guard-Gen-0.6B labelled 275 of 288 unsafe against 254, and the 8B judge 269 against 247, both beyond noise.

Its median thinking on harmful prompts is about 1.3 times the original's. huihui-ai's is 2.8 times, and its thinking is longer than this model's on 261 of the 288 prompts, which is beyond noise.

All 73 edited matrices keep at least 99.3% of the edit at Q4_K_M, so the file is plain Q4_K_M.

Other quantisations haven't been measured, so there's only this one.

Where it comes from

1. Base: Qwen/Qwen3.8-27B, Qwen's original 16-bit weights, with no fine-tune. Text only: the vision encoder isn't included.
2. The edit: a rank-3 change to the attention output matrices in layers 26 to 63 and the MLP output matrices in layers 22 to 56 (73 of the model's 128), removing from each layer the direction it uses to refuse, read where the thinking starts. A 600-trial search on refusals and KL divergence found it as its best edit of this kind. It scored each trial with thinking off, which a check showed ranks edits in nearly the same order as thinking on, then scored its best 6 with thinking on. Its best 4 were then measured in full, and this one was picked as described above.

Use

With llama.cpp, at Qwen's thinking settings and the effort and budget it was measured with:

llama-server -hf candor-lab/Qwen3.8-27B-Candor-Uncensored-GGUF:Q4_K_M -c 32768 \
  --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0 \
  --chat-template-kwargs '{"reasoning_effort": "medium"}' \
  --reasoning-budget 8192 \
  --reasoning-budget-message $'\n\nConsidering the limited time by the user, I have to give the solution based on the thinking directly now.'

Tested in llama-server only. Other GGUF apps, such as LM Studio and Ollama, should work but haven't been checked.

Limits

• It answers requests the original model refuses. You are responsible for how you use it.
• Refusals and maths were measured only in English, with thinking on at reasoning effort medium. An app that doesn't pass reasoning_effort gets the chat template's default, xhigh, which wasn't measured.
• It thinks before it answers, sometimes at length: give it a thinking budget.
• Tool use and agentic tasks weren't measured.
• It's a 27B model, so anything it states as fact can be wrong.

Contact

Requests and feedback are welcome: open a discussion in the Community tab. We read everything but can't promise every request.

Licence

This model is under the Apache License 2.0, in LICENSE, as Qwen3.8-27B is. Qwen3.8-27B is Copyright 2026 Alibaba Cloud. Its weights were changed by the edit described above.