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

Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced

HF 源站

HauhauCS/Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced is an uncensored open-source model hosted on Hugging Face.

Image-Text-to-TextGGUFEnglishLicense: gemma12Buncensoredgemma4visionmultimodalagenticcodingcreative-writingroleplayrpconversational
README.md说明 / 使用方法

license: gemma tags:

  • uncensored
  • gemma4
  • gguf
  • vision
  • multimodal
  • agentic
  • coding
  • creative-writing
  • roleplay
  • rp
  • conversational language:
  • en
  • zh pipeline_tag: image-text-to-text base_model: google/gemma-4-12B-it

Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced

加入 Discord 获取更新、路线图、项目动态,或进行交流互动。

由 HauhauCS 打造的无审查版 Gemma4-12B (QAT)。0/465 拒绝率*

关于模型

数据集或模型能力未作任何削减——功能完整,100% 还原原作者预期,仅仅移除了安全拒绝机制。基于官方 QAT(量化感知训练)权重构建,因此 4-bit 量化版本的质量非常接近全精度水准。

Balanced 平衡版

Balanced 变体(推荐——99% 以上的用户使用该版本即可获得满意体验)采用了优化的全面无审查微调,特别针对智能体编码(agentic coding)、逻辑推理、创意写作以及对可靠性要求极高的任务进行了调优。该模型在作答前会先进行思考推理,表现稳定可靠且严格遵循指令。至于用于 Balanced 仍存在过多回避倾向场景的 Aggressive(激进)变体,根据当前测试来看并无必要。

配合 MTP 提速约 60%

附带用于**推测解码(speculative decoding)**的 MTP(multi-token-prediction,多标记预测)草稿头(draft head)——在保持输出完全一致的前提下,生成速度提升约 60%(模型会验证每一个预测的草稿 token,因此输出质量不受任何影响——带来纯粹的速度提升)。本版本经过专门调优,能够与附带的 MTP 头完美协同工作。

llama.cpp 启动示例:

llama-server \
  -m Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced-Q4_K_M.gguf \
  -md mtp-gemma-4-12B-it.gguf --spec-type draft-mtp \
  -ngl 99 -fa on

注意: MTP 加速效果目前由本人在 llama.cpp(llama-server / llama-cli)环境下测试验证。

文件下载

文件名类型大小
Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced-Q4_K_M.ggufQ4_K_M(文本)6.9 GB
mmproj-Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced-BF16.ggufmmproj(视觉)168 MB
mtp-gemma-4-12B-it.ggufMTP 推测草稿模型242 MB

为什么只提供 Q4_K_M? Gemma 4 针对 ~4-bit 进行了量化感知训练(QAT),因此 Q4_K_M 是最佳平衡点——更高精度的量化只会增加文件体积,却无法带来实际的质量提升。精心量化,专为 4-bit 下的最佳质量打造。

视觉多模态能力

加载模型时同时挂载 mmproj 即可支持图像输入:

llama-server -m Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced-Q4_K_M.gguf \
  --mmproj mmproj-Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced-BF16.gguf -ngl 99 -fa on

推荐采样参数

这些参数是专为本 HauhauCS 构建版本调优的——建议使用这些参数以获得预期的表现和输出质量:

  • temperature 0.6
  • top_k 64
  • top_p 0.9
  • min_p 0.05
  • repeat_penalty 1.1

本发布版进行了端到端的独立调优;上述设置是该调优的一部分,并非 Gemma 原生默认参数。

规格参数

  • 12B Dense(稠密模型)· 256K (262144) 上下文长度
  • 通过 mmproj 支持视觉能力(图像输入)
  • 基于 Google DeepMind 的 Gemma 4 12B

兼容性

  • 支持 llama.cpp、LM Studio、Jan、koboldcpp 及其他 GGUF 运行时环境。
  • 多 GPU + LM Studio: 我个人注意到 Gemma 4 在 LM Studio 的 tensor-split(张量切分)模式下可能会崩溃——对于该模型,建议使用单 GPU(或采用 layer-split 层切分 / 优先级分配方式)。

致谢

  • Google DeepMind — Gemma 4。
  • 包含的 mtp-gemma-4-12B-it.gguf 推测解码草稿头来自 Unsloth 团队的 Gemma 4 发布版本——非常感谢 Unsloth 团队的贡献。

* 经过自动化基准测试和人工拒绝基准测试——在常规使用中未发现拒绝情况。极少数边缘情况提示词在首次提问时可能会发生偏转回避,但重新提问或采用策略性提示框架后即可正常作答。如果您遇到确实阻碍实际使用场景的案例,请加入 Discord 反馈,以便我在未来版本中进行改进优化。

源站原文

license: gemma tags:

  • uncensored
  • gemma4
  • gguf
  • vision
  • multimodal
  • agentic
  • coding
  • creative-writing
  • roleplay
  • rp
  • conversational language:
  • en pipeline_tag: image-text-to-text base_model: google/gemma-4-12B-it

Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced

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

Gemma4-12B (QAT) uncensored by HauhauCS. 0/465 Refusals*

About

No changes to datasets or capabilities — fully functional, 100% of what the original authors intended, just without the refusals. Built from the official QAT weights, so the 4-bit quant stays close to full-precision quality.

Balanced

The Balanced variant (recommended — 99%+ of users will be happy here) uses optimized full uncensoring tuned especially for agentic coding, reasoning, creative writing and reliability-critical tasks. It reasons before answering and stays dependable and on-instruction. An Aggressive variant, for cases where Balanced still deflects too much, after current testing is not required.

~60% faster with MTP

Ships with an MTP (multi-token-prediction) draft head for speculative decoding — roughly 60% faster generation with identical output (the model verifies every drafted token, so quality is unchanged — pure speed). This release is tuned to pair well with the included MTP head.

llama.cpp:

llama-server \
  -m Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced-Q4_K_M.gguf \
  -md mtp-gemma-4-12B-it.gguf --spec-type draft-mtp \
  -ngl 99 -fa on

Note: the MTP speedup was currently tested by me through llama.cpp (llama-server / llama-cli).

Downloads

FileTypeSize
Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced-Q4_K_M.ggufQ4_K_M (text)6.9 GB
mmproj-Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced-BF16.ggufmmproj (vision)168 MB
mtp-gemma-4-12B-it.ggufMTP speculative drafter242 MB

Why only Q4_K_M? Gemma 4 is quantization-aware-trained for ~4-bit, so Q4_K_M is the sweet spot — higher-precision quants add size with no real quality gain. Carefully quantized for best quality at 4-bit.

Vision

Load the mmproj alongside the model for image input:

llama-server -m Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced-Q4_K_M.gguf \
  --mmproj mmproj-Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced-BF16.gguf -ngl 99 -fa on

Recommended sampling

These are dialed in specifically for this HauhauCS build — use them for the intended behaviour and quality:

  • temperature 0.6
  • top_k 64
  • top_p 0.9
  • min_p 0.05
  • repeat_penalty 1.1

This release is tuned end-to-end as its own thing; the settings above are part of that and aren't the stock Gemma defaults.

Specs

  • 12B dense · 256K (262144) context
  • Vision (image input) via mmproj
  • Based on Gemma 4 12B by Google DeepMind

Compatibility

  • Works with llama.cpp, LM Studio, Jan, koboldcpp, and other GGUF runtimes.
  • Multi-GPU + LM Studio: I've personally noticed Gemma 4 can crash under LM Studio's tensor-split mode — use a single GPU (layer-split or priority order) for this model.

Acknowledgements

  • Google DeepMind — Gemma 4.
  • The included mtp-gemma-4-12B-it.gguf speculative draft head comes from Unsloth's Gemma 4 release — many thanks to the Unsloth team for it.

* Tested with both automated and manual refusal benchmarks — none have been found in standard use. A small number of edge-case prompts deflect on the first ask but comply on a re-ask or strategic framing. If you hit one that's actually obstructive to your use case, join the Discord and flag it so I can work on it in a future revision.

HUGGING FACE

上游数据

HF 下载量64.8万
HF 点赞量353

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
2026/06/25
最新创建
2026/06/22
HF 趋势
38
数据快照
2026/09/16
版本
未知
许可证
gemma