玩模型开放模型 · 清晰来源 · 自主选择
返回模型目录
HauhauCS /

Gemma-4-E2B-Uncensored-HauhauCS-Aggressive

HF 源站

HauhauCS/Gemma-4-E2B-Uncensored-HauhauCS-Aggressive is an uncensored open-source model hosted on Hugging Face.

Image-Text-to-TextGGUFEnglishMultilingualLicense: gemmauncensoredgemma4visionmultimodalaudioabliteratedimatrixconversational
README.md说明 / 使用方法

license: gemma tags:

  • uncensored
  • gemma4
  • gguf
  • vision
  • multimodal
  • audio
  • abliterated language:
  • en
  • multilingual pipeline_tag: image-text-to-text base_model: google/gemma-4-e2b-it

Gemma-4-E2B-Uncensored-HauhauCS-Aggressive

加入 Discord 获取更新动态、开发路线图、项目进展或参与日常交流。

由 HauhauCS 进行无审查(uncensored)处理的 Gemma 4 E2B-IT。0/465 次拒绝(Refusals)***

HuggingFace 的“Hardware Compatibility”(硬件兼容性)组件无法识别 K_P 量化版本 —— 它显示的文件数量可能少于实际存在的文件。请点击 “View +X variants” 或前往 Files and versions 查看所有可供下载的文件。

关于本项目

未对数据集或模型能力做任何改动。功能完全正常,100% 保留了原作者预期的全部能力 —— 只是去除了拒绝回答的行为。

本系列旨在打造目前市面上最好的无损无审查(uncensored)模型。

激进版本(Aggressive Variant)

更强力的去审查处理 —— 模型完全解锁,不会拒绝提示词。可能会偶尔附加简短的免责声明(这是固化在基座模型训练中的,并非拒绝回答),但完整内容始终会被生成。

如果您希望获得更保守的去审查效果并保留部分安全防护栏,可在可用时查看 Balanced(平衡)版本。

模型下载

文件量化版本BPW大小
Gemma-4-E2B-Uncensored-HauhauCS-Aggressive-Q8_K_P.ggufQ8_K_P9.44.7 GB
—Q8_08.5—
Gemma-4-E2B-Uncensored-HauhauCS-Aggressive-Q6_K_P.ggufQ6_K_P7.03.7 GB
—Q6_K6.6—
Gemma-4-E2B-Uncensored-HauhauCS-Aggressive-Q5_K_P.ggufQ5_K_P6.13.5 GB
Gemma-4-E2B-Uncensored-HauhauCS-Aggressive-Q4_K_P.ggufQ4_K_P5.23.3 GB
Gemma-4-E2B-Uncensored-HauhauCS-Aggressive-Q3_K_P.ggufQ3_K_P4.13.1 GB
Gemma-4-E2B-Uncensored-HauhauCS-Aggressive-IQ3_M.ggufIQ3_M3.73.0 GB
Gemma-4-E2B-Uncensored-HauhauCS-Aggressive-Q2_K_P.ggufQ2_K_P3.52.9 GB
mmproj-Gemma-4-E2B-Uncensored-HauhauCS-Aggressive-f16.ggufmmproj (f16)—940 MB

所有量化版本均使用重要性矩阵(imatrix)生成,以在消除审查(abliterated)的权重上实现最佳质量保留。

什么是 K_P 量化?

K_P(“Perfect”,完美)量化是 HauhauCS 的自定义量化方法,它利用针对特定模型的分析,选择性地在最关键的部分保留精度质量。每个模型都会获得其专属的优化量化配置方案。

相比基础量化版本,K_P 量化仅增加约 5-15% 的文件体积,即可有效将模型输出质量提升 1 到 2 个量化档位。完全兼容 llama.cpp、LM Studio 以及任何支持 GGUF 的运行时环境 —— 无需特殊构建。

注意: K_P 量化在 LM Studio 的量化列中可能会显示为“?”。这仅仅是一个界面显示问题 —— 模型可以正常加载并运行。

规格参数

  • 2B 参数量
  • 35 层,混合滑动窗口(512)+ 全注意力机制(full attention)
  • 131K 上下文长度
  • 原生多模态(文本、图像、视频、音频)
  • 20 个 KV 共享层以提升显存效率
  • 基于 google/gemma-4-e2b-it

推荐运行配置

来自 Google Gemma 4 官方作者推荐:

  • temperature=1.0, top_p=0.95, top_k=64

重要提示:

  • 使用 llama.cpp 时请带上 --jinja 参数,以确保正确解析聊天模板(chat template)
  • 视觉/音频功能需要搭配主 GGUF 对应的 mmproj 文件共同使用

使用方法

兼容 llama.cpp、LM Studio、Jan、koboldcpp 及其他支持 GGUF 的运行时。

# 仅文本
llama-cli -m Gemma-4-E2B-Uncensored-HauhauCS-Aggressive-Q4_K_P.gguf \
  --jinja -c 8192 -ngl 99

# 启用视觉/音频
llama-cli -m Gemma-4-E2B-Uncensored-HauhauCS-Aggressive-Q4_K_P.gguf \
  --mmproj mmproj-Gemma-4-E2B-Uncensored-HauhauCS-Aggressive-f16.gguf \
  --jinja -c 8192 -ngl 99

其他尺寸

  • Gemma-4-E4B-Uncensored-HauhauCS-Aggressive — 4B 版本,性能更强

* 相比我的其他发布版本,Gemma 4 在长上下文场景下没有经过同样长时间的人工测试。Google 目前正在使用类似 NVIDIA GenRM 的技术 —— 即作为内部批判器的生成式奖励模型(generative reward models)—— 这使得(真正意义上的)去审查变成了一个愈发具有挑战性的领域。我认为 99.999% 的用户不会遇到边界极端情况,但出于诚实负责的考虑,依然在此添加星号说明。

** 这是一个 2B 规模的模型。请合理设定预期 —— 以其尺寸而言它确实令人印象深刻,但它毕竟只有 20 亿参数。复杂推理、细腻角色扮演以及长篇连贯生成并非其强项。非常适合轻量快速任务、移动端/边缘端部署以及实验性探索。

源站原文

license: gemma tags:

  • uncensored
  • gemma4
  • gguf
  • vision
  • multimodal
  • audio
  • abliterated language:
  • en
  • multilingual pipeline_tag: image-text-to-text base_model: google/gemma-4-e2b-it

Gemma-4-E2B-Uncensored-HauhauCS-Aggressive

Join the Discord for updates, roadmaps, projects, or just to chat.

Gemma 4 E2B-IT uncensored by HauhauCS. 0/465 Refusals***

HuggingFace's "Hardware Compatibility" widget doesn't recognize K_P quants — it may show fewer files than actually exist. Click "View +X variants" or go to Files and versions to see all available downloads.

About

No changes to datasets or capabilities. Fully functional, 100% of what the original authors intended - just without the refusals.

These are meant to be the best lossless uncensored models out there.

Aggressive Variant

Stronger uncensoring — model is fully unlocked and won't refuse prompts. May occasionally append short disclaimers (baked into base model training, not refusals) but full content is always generated.

For a more conservative uncensor that keeps some safety guardrails, check the Balanced variant when it's available.

Downloads

FileQuantBPWSize
Gemma-4-E2B-Uncensored-HauhauCS-Aggressive-Q8_K_P.ggufQ8_K_P9.44.7 GB
—Q8_08.5—
Gemma-4-E2B-Uncensored-HauhauCS-Aggressive-Q6_K_P.ggufQ6_K_P7.03.7 GB
—Q6_K6.6—
Gemma-4-E2B-Uncensored-HauhauCS-Aggressive-Q5_K_P.ggufQ5_K_P6.13.5 GB
Gemma-4-E2B-Uncensored-HauhauCS-Aggressive-Q4_K_P.ggufQ4_K_P5.23.3 GB
Gemma-4-E2B-Uncensored-HauhauCS-Aggressive-Q3_K_P.ggufQ3_K_P4.13.1 GB
Gemma-4-E2B-Uncensored-HauhauCS-Aggressive-IQ3_M.ggufIQ3_M3.73.0 GB
Gemma-4-E2B-Uncensored-HauhauCS-Aggressive-Q2_K_P.ggufQ2_K_P3.52.9 GB
mmproj-Gemma-4-E2B-Uncensored-HauhauCS-Aggressive-f16.ggufmmproj (f16)—940 MB

All quants generated with importance matrix (imatrix) for optimal quality preservation on abliterated weights.

What are K_P quants?

K_P ("Perfect") quants are HauhauCS custom quantizations that use model-specific analysis to selectively preserve quality where it matters most. Each model gets its own optimized quantization profile.

A K_P quant effectively bumps quality up by 1-2 quant levels at only ~5-15% larger file size than the base quant. Fully compatible with llama.cpp, LM Studio, and any GGUF-compatible runtime — no special builds needed.

Note: K_P quants may show as "?" in LM Studio's quant column. This is a display issue only — the model loads and runs fine.

Specs

  • 2B parameters
  • 35 layers, mixed sliding window (512) + full attention
  • 131K context
  • Natively multimodal (text, image, video, audio)
  • 20 KV shared layers for memory efficiency
  • Based on google/gemma-4-e2b-it

Recommended Settings

From the official Google Gemma 4 authors:

  • temperature=1.0, top_p=0.95, top_k=64

Important:

  • Use --jinja flag with llama.cpp for proper chat template handling
  • Vision/audio support requires the mmproj file alongside the main GGUF

Usage

Works with llama.cpp, LM Studio, Jan, koboldcpp, and other GGUF-compatible runtimes.

# Text only
llama-cli -m Gemma-4-E2B-Uncensored-HauhauCS-Aggressive-Q4_K_P.gguf \
  --jinja -c 8192 -ngl 99

# With vision/audio
llama-cli -m Gemma-4-E2B-Uncensored-HauhauCS-Aggressive-Q4_K_P.gguf \
  --mmproj mmproj-Gemma-4-E2B-Uncensored-HauhauCS-Aggressive-f16.gguf \
  --jinja -c 8192 -ngl 99

Other Sizes

  • Gemma-4-E4B-Uncensored-HauhauCS-Aggressive — 4B version, more capable

* Gemma 4 didn't get as much manual testing time at longer context as my other releases. Google is now using techniques similar to NVIDIA's GenRM — generative reward models that act as internal critics — making (true) uncensoring an increasingly challenging field. I expect 99.999% of users won't hit edge cases, but the asterisk is there for honesty.

** This is a 2B model. Temper your expectations — it's impressive for its size, but it's still 2B parameters. Complex reasoning, nuanced roleplay, and long coherent outputs are not its strong suit. Great for quick tasks, mobile/edge deployment, and experimentation.

HUGGING FACE

上游数据

HF 下载量13.2万
HF 点赞量312

HF 下载量为源站最近 30 天统计,与本站点击分开。

SITE STATS

本站统计

下载点击—
请求数—
匿名去重—最近 7 天—

本站下载量记录按钮点击,不代表下载完成;短时连点会去重。

来源与版本

作者
HauhauCS
最近更新
2026/04/06
最新创建
2026/04/02
HF 趋势
94
数据快照
2026/09/16
版本
未知
许可证
gemma