Modelpublic

Suseezz/Qwen3.8-27B-Uncensored-IQ4-XS-MTP-16GB-VRAM-GGUF

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

0likes375downloads

Model Card

此模型可能短期更新，我发布了一个基于Heretic Arbitrary-Rank Ablation的无审查版本模型，性能更好且体积更小链接： https://huggingface.co/Bucoid/Qwen3.8-27B-Heretic-Ara-IQ4-XS-16GB-VRAM-GGUF 这个模型可能过一段时间我会更新让他不那么菜，如果你需要无审查版本的模型，基于下载这个Ara的

This model may receive short-term updates. I have released an uncensored version based on Heretic Arbitrary-Rank Ablation, which offers better performance and a smaller file size. Link: https://huggingface.co/Bucoid/Qwen3.8-27B-Heretic-Ara-IQ4-XS-16GB-VRAM-GGUF This model may be updated in a while to make it less underwhelming If you need an uncensored version, please download this Ara-based one instead.

Qwen3.8-27B Uncensored IQ4XS 量化模型（适配 16GB 显存）本模型基于 Qwen3.8-27B Uncensored 进行 IQ4XS 量化（4‑bit），文件体积为 12.9 GiB，专为 16GB 显存的显卡优化，在保持较低困惑度的同时，兼顾推理速度和显存占用。

与同体积的 UDIQ3K_XL（12.5 GiB）量化方案进行了全面对比，评估指标如下。

📊 量化质量对比

评估指标	IQ4_XS (本模型)	UD_IQ3_K_XL (对比)
文件大小	12.9 GB	12.5 GB
量化精度	IQ4_XS (4‑bit)	UDIQ3K_XL (约 3‑bit?)
量化模型困惑度 (Mean PPL)	7.1481 ± 0.0465	7.1117 ± 0.0459
与基座模型 PPL 相关性	99.28%	99.31%
平均 KL 散度 (Mean KLD)	0.03268 ± 0.00030	0.03130 ± 0.00032
最大 KL 散度 (Max KLD)	16.017（更小）	21.409
99.9% KL 分位数	1.075	1.219
Top‑1 一致率 (Same top p)	91.655% ± 0.072%	92.419% ± 0.069%
平均概率变化 (Mean Δp)	-0.343% ± 0.013%（更接近 0）	-0.738% ± 0.013%
RMS 概率变化 (RMS Δp)	4.986% ± 0.039%（更小）	5.120% ± 0.046%

在不启用MTP的情况下可以做到16GiB净空VRAM（不作为Windows的显示显卡）的情况下110k上下文

开启MTP大概80k上下文。

license: apache-2.0 base_model:

—Qwen/Qwen3.8-27B --- Qwen3.8-27B Uncensored IQ4XS Quantized Model (Optimized for 16GB VRAM) This model is based on Qwen3.8-27B Uncensored and quantized with IQ4XS (4‑bit), with a file size of 12.9 GiB. It is tailored for GPUs with 16GB VRAM, balancing low perplexity, inference speed, and memory usage.

We conducted a comprehensive comparison against the UDIQ3K_XL quantization scheme (12.5 GiB, roughly 3‑bit) of the same model size. The evaluation metrics are as follows.

📊 Quantization Quality Comparison

Metric	IQ4_XS (this model)	UD_IQ3_K_XL (baseline)
File size	12.9 GB	12.5 GB
Quantization precision	IQ4_XS (4‑bit)	UDIQ3K_XL (~3‑bit)
Mean perplexity (quantized)	7.1481 ± 0.0465	7.1117 ± 0.0459
Correlation with base model PPL	99.28%	99.31%
Mean KL divergence	0.03268 ± 0.00030	0.03130 ± 0.00032
Maximum KL divergence	16.017 (lower)	21.409
99.9% KL quantile	1.075	1.219
Top‑1 agreement rate	91.655% ± 0.072%	92.419% ± 0.069%
Mean probability change (Mean Δp)	-0.343% ± 0.013% (closer to 0)	-0.738% ± 0.013%
RMS probability change (RMS Δp)	4.986% ± 0.039% (lower)	5.120% ± 0.046%

With MTP disabled, the model can achieve ~110k context length while keeping ~16 GiB free VRAM (when not used as the primary display GPU on Windows). With MTP enabled, the context length is around 80k.