Loadingβ¦
Loadingβ¦
Qwen/Qwen3.8-27BDownload the file and run it. Delete the file when you are done.
No environment setup. No library installation. No annoying clean up.
Compare uncensored variants β
These bundled files are too large to run on Windows. Use a separate llamafile binary with the source GGUF files, as shown in the llamafile docs.
Download the llamafile (e.g. Q4_K_M).
Open a terminal on Linux or macOS.
Grant execute permission.
chmod +x Qwen3.8-27B-OBLITERATED-Q4_K_M.llamafile
Run the file.
bash Qwen3.8-27B-OBLITERATED-Q4_K_M.llamafile
# Or run directly if your system supports it.
./Qwen3.8-27B-OBLITERATED-Q4_K_M.llamafile
http://127.0.0.1:8080/v1./v1/chat/completions.Press Control-C to stop the process.
For more details, run with --help or read the llamafile docs.
bash Qwen3.8-27B-OBLITERATED-Q4_K_M.llamafile --help
ubuntu@ubuntu > ./Qwen3.8-27B-OBLITERATED-Q4_K_M.llamafile
βββ βββ ββββββ ββββ ββββ ββββββ ββββββββββββββ ββββββββ
βββ βββ βββββββββββββ βββββββββββββββββββββββββββ ββββββββ
βββ βββ βββββββββββββββββββββββββββββββββ ββββββ ββββββ
βββ βββ βββββββββββββββββββββββββββββββββ ββββββ ββββββ
βββββββββββββββββββ ββββββ βββ ββββββ ββββββ βββββββββββββββββββ
βββββββββββββββββββ ββββββ ββββββ ββββββ βββββββββββββββββββ
software: llamafile 0.10.5
model: Qwen3.8-27B-OBLITERATED-Q4_K_M.gguf
compute: Apple Metal GPU
server: http://127.0.0.1:8080
A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.
>>>
Part of Qwen3.8-27B Uncensored Variant Llamafiles.
Chart source code#@title Performance versus uncensored response rate
import matplotlib.pyplot as plt
from statistics import mean
# Source: [Abliterlitics](https://web.archive.org/web/20260909182157/https://abliterlitics.dev/models/qwen38-27b/)
# Archived FP8 results, not measurements of these GGUFs.
# Columns: MMLU-Pro, GPQA, BBH, MuSR, IFEval, HellaSwag, ARC-C, WinoGrande,
# PIQA, GSM8K, TQA-MC1, TQA-MC2, TQA-Gen, HumanEval, HarmBench judge ASR (%).
data = {
"Official": [62.82, 28.28, 75.16, 45.90, 82.07, 74.79, 59.90, 71.11, 79.65, 91.21, 44.43, 66.98, 55.08, 79.3, 4.5],
"Heretic ARA": [62.65, 27.27, 74.78, 45.90, 82.26, 74.19, 60.24, 71.27, 79.65, 91.51, 44.19, 65.74, 55.08, 82.3, 57.5],
"Apostate": [62.81, 29.80, 74.22, 45.63, 82.99, 74.54, 60.84, 71.19, 79.87, 91.81, 44.19, 65.02, 53.37, 79.9, 78.7],
"coder3101": [62.66, 28.79, 74.78, 46.16, 83.36, 74.53, 59.90, 70.09, 79.60, 91.21, 43.57, 64.63, 53.12, 80.5, 70.0],
"OrcaRouter": [62.78, 27.27, 74.29, 45.37, 82.44, 74.16, 59.90, 71.27, 79.65, 90.75, 44.55, 65.48, 53.49, 78.7, 82.2],
"Ultra Heretic": [62.25, 26.77, 74.92, 45.24, 82.26, 74.34, 58.28, 71.11, 79.22, 90.75, 39.90, 60.54, 48.84, 79.3, 70.5],
"Huihui": [62.67, 32.32, 74.69, 45.63, 81.89, 74.10, 60.07, 71.35, 79.43, 91.74, 41.86, 63.71, 52.02, 78.7, 75.6],
"OBLITERATUS": [56.49, 30.30, 70.44, 45.90, 77.63, 73.90, 62.03, 69.93, 80.69, 92.19, 38.19, 55.82, 45.04, 45.7, 63.9],
"Blackfrost": [60.89, 25.76, 73.36, 42.72, 82.44, 73.31, 58.19, 68.11, 79.82, 89.76, 43.45, 64.08, 51.04, 70.7, 68.5],
}
# Not in this collection; remove these filters to include all source rows.
del data["Apostate"]
del data["coder3101"]
# Equal mean of 12 families; average TQA's three scores first. Exclude perplexity.
performance = [mean(row[:10] + [mean(row[10:13]), row[13]]) for row in data.values()]
# Blog judge ASR excludes degenerate responses; it is not final-answer usability.
uncensored = [row[-1] for row in data.values()]
fig, ax = plt.subplots(figsize=(9, 6), layout="constrained")
ax.scatter(performance, uncensored, s=55)
for name, x, y in zip(data, performance, uncensored):
ax.annotate(name, (x, y), xytext=(6, 6), textcoords="offset points")
ax.set(xlabel="Performance β mean of 12 benchmark families (%)",
ylabel="Uncensored response rate β HarmBench judge ASR (%)",
title="Qwen3.8-27B variants", ylim=(0, 100))
ax.margins(x=0.20)
ax.grid(alpha=0.2)
fig.savefig("qwen38-performance-vs-asr.png", dpi=180, bbox_inches="tight")
plt.show()
Other variants:
The suggestions are based on imperfect benchmarks and should not be trusted blindly. Try the variants yourself with your own use case.
All data came from the Abliterlitics review (snapshot), which also contains more detailed analysis of more uncensored model variants. I am not the author, nor did I replicate the experiments myself.
a58c3b53b3ce71551eafde2ed5ec8df48e0f4ff8486e6c5f9356eae50b851b07517bfae1f2420193validation/..args contents:-m
Qwen3.8-27B-OBLITERATED-Q4_K_M.gguf
...