返回模型目录Jiunsong /
supergemma4-26b-uncensored-gguf-v2
Jiunsong/supergemma4-26b-uncensored-gguf-v2 is an uncensored open-source model hosted on Hugging Face.
license: gemma
base_model: google/gemma-4-26B-A4B-it
tags:
- gguf
- gemma4
- uncensored
- fast
- llama.cpp
- apple-silicon
- conversational
- korean
- coding
- tool-use
language:
- en
- ko
pipeline_tag: text-generation
SuperGemma4-26B-Uncensored-Fast GGUF v2
这是基于最强 SuperGemma 文本模型系列构建的高速、未审查（uncensored）llama.cpp 版本。
本发布版本专为同时追求以下三点的用户打造：
- 相比原生官方对话模型更少审查限制的模型体验
- 在实际文本工作负载上比原始基座模型性能更强大
- 针对 Apple Silicon 优化、保持紧凑且能快速推理的本地 GGUF 模型
为何选择此版本
- 具备未审查的对话交互特性，且不会将每个提示词都强行转向编程模式
- 基于性能最强的
fast 系列进行微调，而非原始基座模型
- GGUF 内部内置中立的对话模板（chat template），大幅减少提示词路由错误（prompt-routing bugs）
- 在 Apple Silicon 上经过验证，通用对话与代码生成均表现优异
核心指标
- 基座模型：
google/gemma-4-26B-A4B-it
- 格式：
GGUF Q4_K_M
- 韩语通用提示处理速度：
222.0 tok/s
- 生成速度：
89.4 tok/s
- 衍生自经过验证的
SuperGemma Fast MLX 系列
本版本的核心优势
- 继承了更强大的
Fast 权重，优于普通官方基座
- 保持通用对话自然流畅，不会将所有问题都引导至编程模式
- 既保留了未审查版本的个性特色，又能在常规提示交互中保持实用性
- 提供了实用的
llama.cpp 部署目标，同时不丢失微调系列固有的独特个性
为何优于原生版本
- 继承了
Fast 系列相对于原始本地基准的显著改进：
- 快速基准测试（Quick bench）综合得分：
95.8 vs 91.4
- MLX 参考运行中的平均生成速度更快：
46.2 tok/s vs 42.5 tok/s
- 在代码、逻辑推理、浏览器工作流及韩语能力上得分更高
- 内置中立嵌入式模板，避免了此前简单问题被偏向引导至代码编写/工具调用行为的路由问题
包含文件
supergemma4-26b-uncensored-fast-v2-Q4_K_M.gguf
本地快速实测
在配备 llama.cpp 的 Apple M4 Max 上测试：
- 韩语通用提示：
봄에 먹기 좋은 한식 반찬 5개 추천
- 提示处理速度：
222.0 tok/s
- 生成速度：
89.4 tok/s
- 输出稳定保持在正常的韩语助手模式
- 代码提示：
파이썬으로 피보나치 함수를 짧게 작성해줘
- 提示处理速度：
704.9 tok/s
- 生成速度：
89.4 tok/s
- 正确返回了简洁的 Python 代码
注意事项
- 此 GGUF 导出自
supergemma4-26b-uncensored-fast-v2 MLX 系列。
- Gemma 4 MoE 专家张量（expert tensors）采用打过补丁的本地转换器进行转换，确保 GGUF 导出正常运行。
- 嵌入了中立模板，以避免通用提示词被强制转换为代码编写/工具调用的旧问题。
源站原文
license: gemma
base_model: google/gemma-4-26B-A4B-it
tags:
- gguf
- gemma4
- uncensored
- fast
- llama.cpp
- apple-silicon
- conversational
- korean
- coding
- tool-use
language:
- en
- ko
pipeline_tag: text-generation
SuperGemma4-26B-Uncensored-Fast GGUF v2
The fast, uncensored llama.cpp build of the strongest SuperGemma text line.
This release is for people who want three things together:
- a model that feels less censored than stock chat releases
- a model that is more capable than the raw base on practical text workloads
- a compact local GGUF that still serves quickly on Apple Silicon
Why this build
- Uncensored chat behavior without forcing every prompt into coding mode
- Tuned from the strongest
fast line instead of the raw base
- Neutral chat template baked into the GGUF to reduce prompt-routing bugs
- Verified on Apple Silicon with clean general-chat and coding responses
Headline numbers
- Base model:
google/gemma-4-26B-A4B-it
- Format:
GGUF Q4_K_M
- General Korean prompt speed:
222.0 tok/s
- Generation speed:
89.4 tok/s
- Derived from the verified
SuperGemma Fast MLX line
Why this build is appealing
- Carries the stronger
Fast weights instead of the plain stock base
- Keeps general chat natural instead of routing everything into coding mode
- Preserves the uncensored release identity while staying useful on normal prompts
- Gives you a practical
llama.cpp deployment target without losing the personality of the tuned line
Why it is better than stock
- Inherits the
Fast line improvements over the original local baseline:
- Quick bench overall:
95.8 vs 91.4
- Faster average generation on the MLX reference run:
46.2 tok/s vs 42.5 tok/s
- Higher scores in code, logic, browser workflows, and Korean
- Ships with a neutral embedded template to avoid the older routing bug where simple questions drifted into coding/tool-call behavior
Included file
supergemma4-26b-uncensored-fast-v2-Q4_K_M.gguf
Quick local checks
Tested on Apple M4 Max with llama.cpp:
- General Korean prompt:
봄에 먹기 좋은 한식 반찬 5개 추천
- Prompt speed:
222.0 tok/s
- Generation speed:
89.4 tok/s
- Output stayed in normal Korean assistant mode
- Code prompt:
파이썬으로 피보나치 함수를 짧게 작성해줘
- Prompt speed:
704.9 tok/s
- Generation speed:
89.4 tok/s
- Output returned concise Python code correctly
Notes
- This GGUF is exported from the
supergemma4-26b-uncensored-fast-v2 MLX line.
- Gemma 4 MoE expert tensors were converted with a patched local converter so GGUF export works correctly.
- A neutral template is embedded to avoid the old issue where general prompts were pushed into coding/tool-call behavior.
