license: apache-2.0
license_link: LICENSE
base_model: orcarouter/Qwen3.8-27B-Uncensored-GGUF
base_model_relation: quantized
pipeline_tag: text-generation
library_name: gguf
tags:
- qwen
- qwen3.8
- gguf
- llama.cpp
- abliterated
- uncensored
Qwen3.8-27B-Uncensored-GGUF
This repo holds three GGUF files from orcarouter/Qwen3.8-27B-Uncensored-GGUF. That is an abliterated (refusal-removed) build of Qwen/Qwen3.8-27B.
The files are byte-for-byte copies of orcarouter's files. They have not been requantized or modified. They are republished here, under the Apache-2.0 license, as the Silver, Gold and Platinum tiers of DeepSeek Unchained. All credit for the model goes to the Qwen team, and for the abliteration and quantization to orcarouter.
Disclaimer
This model has had its safety alignment substantially removed. It will comply with requests that the original Qwen3.8-27B refuses. You are responsible for how you use it and for what it generates. Use it within the Apache-2.0 license and applicable law. orcarouter's model card has the full disclaimer.
Files
| Path | Quant | Size (bytes) | sha256 |
|---|---|---|---|
IQ3_M/Qwen3.8-27B-Uncensored-IQ3_M.gguf |
IQ3_M | 12,768,331,360 | 045ca1b0396bda8591e8b6f8c616c54ab2e6c4e0aea0275e5107d1eccf6204ec |
Q4_K_M/Qwen3.8-27B-Uncensored-Q4_K_M.gguf |
Q4_K_M | 16,810,714,496 | 3445102e9cde5d562508642c100a2f5ac3368a5a3f748442811d7a95daee3bec |
Q8_0/Qwen3.8-27B-Uncensored-Q8_0.gguf |
Q8_0 | 29,047,084,416 | 31756fca94beca71ea4b8706d6fdc896dab2a3c6376ab0c1863b98512a24f8d6 |
Each hash matches the file in the orcarouter repo. These are the text-only files: no MTP head and no mmproj. A recent llama.cpp with the qwen35 architecture is required. These files were tested with b10919.
How the quants were chosen
Six of orcarouter's quants were measured against its F16 file on one Apple M4 Max with 128 GB of memory. The tool was llama.cpp b10919 llama-perplexity (KL divergence against the F16 logits, wikitext-2-raw test, ctx 512, 100 chunks), and llama-bench was used for speed.
| Quant | File GB | Mean KLD | p99 KLD | Same top token % | PPL ratio vs F16 | pp512 t/s | tg128 t/s |
|---|---|---|---|---|---|---|---|
| IQ3_M | 12.77 | 0.0576 | 0.568 | 89.78 | 1.0348 | 213.0 | 18.21 |
| Q3_K_M | 13.50 | 0.0687 | 0.635 | 88.75 | 1.0472 | 206.7 | 16.26 |
| Q4_K_M | 16.81 | 0.0205 | 0.201 | 93.75 | 1.0039 | 209.8 | 17.71 |
| Q5_K_M | 19.54 | 0.0064 | 0.063 | 96.49 | 1.0045 | 188.2 | 12.93 |
| Q6_K | 22.43 | 0.0023 | 0.020 | 97.82 | 1.0003 | 162.6 | 13.16 |
| Q8_0 | 29.05 | 0.0003 | 0.002 | 99.33 | 1.0014 | 182.3 | 10.93 |
IQ3_M beat Q3_K_M on every measure, so it became the smallest tier. Lower KLD means the quant's output distribution is closer to F16. "Same top token" is how often the quant and F16 agree on the most likely next token.
License
Apache-2.0, inherited from Qwen/Qwen3.8-27B. The LICENSE file is copied from that repo.