Harsha001/Arete-27B-Uncensored-ROCmFPX

Arete-27B Uncensored, ROCmFPX GGUF

Quantized files of SargeDev's uncensored Arete-27B, a 27B text model. The original, Arete-27B by Nous Research, is a values-aligned full fine-tune of Qwen3.8-27B. SargeDev removed its refusal direction with OBLITERATUS and added the stock Qwen3.8 multi-token prediction (MTP) head.

Warning: safety training has been removed from this model. The abliteration removes its refusal behaviour, so it can comply with harmful, unethical, offensive or illegal requests. You are fully responsible for how you use it and for what it generates, and you must follow the law that applies to you. Do not put it in front of end users without your own safety and moderation layers. Its outputs do not reflect the views of the uploader or of the original authors.

These files have not been tested on AMD hardware yet. Please report speed and quality in the Community tab.

The 8-bit and 6-bit files need HarshaVardhan3002/ROCmFPX, built from 7 November 2026 or later. Stock llama.cpp, LM Studio, Ollama and older ROCmFPX builds will not load them. All files here were only tested with that fork.

Not an official release. All credit for the model goes to Nous Research and the Qwen team, and for the uncensoring to SargeDev.

Files

File	Size	KL to BF16	Same top token
Arete-27B-Uncensored-Q8_0_ROCMFPX.gguf	28.3 GB	0.00030	99.2%	8-bit, start here
Arete-27B-Uncensored-Q6_0_ROCMFPX.gguf	22.5 GB	0.0021	97.6%	6-bit
Arete-27B-Uncensored-Q4_K_M.gguf	16.8 GB	0.016	94.7%	4-bit, standard llama.cpp weight type

For comparison, measured the same way:

Not published	Size	KL to BF16	Same top token
Stock Q8_0	29.0 GB	0.00038	99.1%
Stock Q6_K	22.4 GB	0.0024	97.9%
ROCmFPX 4-bit (FP4, attention 8-bit)	19.2 GB	0.021	93.6%

The ROCmFPX 8-bit file is closer to BF16 than a stock Q8_0 and smaller. At 6-bit, ROCmFPX has the lower KL for 0.1 GB more, while the stock Q6_K keeps the same top token slightly more often, so the two are close. At 4-bit the stock Q4_K_M was both smaller and closer, so that is the 4-bit file here.

How this was measured: llama.cpp llama-perplexity --kl-divergence against a BF16 GGUF converted from SargeDev's weights (perplexity 6.28), 40 chunks of 512 tokens of English Wikipedia text (wikitext-2), on an NVIDIA GB10 GPU after expanding each file to F16. KL is how far the file's next-token probabilities drift from BF16, averaged over every token; lower is better. This is general English text, so it says nothing direct about reasoning or the benchmarks on the original card.

Short checks with greedy decoding:

• Text: all three files gave 17 x 23 = 391 and Canberra as Australia's capital, and the 8-bit and 6-bit files answered "What is the capital of France?" correctly after the thinking block.
• Refusals: 1 of 10 requests that a safety-trained model would normally refuse, and 0 of 8 harmless questions that only sound risky, on all three files and on the BF16 source. The one case is the same everywhere: asked how to shoplift, the model first says it cannot help plan a crime and then answers anyway. So this comes from the uncensored model itself, not from quantization. This small test looks for refusal phrases in the first 64 tokens. It is not a measure of how useful or accurate the answers are. (The 8-bit file was tested on CPU; the BF16 source, the Q4_K_M and an F16 expansion of the 6-bit file on the GPU.)

The files contain the one-layer MTP head that SargeDev added, which was not tested.

Credits and license

• Original model: NousResearch/Arete-27B, a fine-tune of Qwen3.8-27B by the Qwen team, Apache-2.0.
• Uncensored version: SargeDev/Arete-27B-Uncensored, made with OBLITERATUS.

Apache-2.0. See LICENSE.