---

license: apache-2.0

base_model: gjtgjt/Muse-Glimmer-30B-heretic

pipeline_tag: image-text-to-text

language:

- zh

- en

tags:

- gguf

- llama.cpp

- heretic

- abliterated

- uncensored

- muse-glimmer

- multimodal

---

Muse-Glimmer-30B-heretic-GGUF

中文 | English

---

中文

这是 gjtgjt/Muse-Glimmer-30B-heretic 的 GGUF 量化版本。该模型由 Meta 的 meta-models/Muse-Glimmer-30B 经 Heretic 消融（abliteration / 去审查）得到。

基础模型许可证：Apache-2.0。

文件

| 文件 | 量化 | 大小 |

|---|---|---|

| Muse-Glimmer-30B-heretic-Q5_K_M.gguf | Q5_K_M | 18.4 GiB |

| Muse-Glimmer-30B-heretic-Q8_0.gguf | Q8_0 | 27.6 GiB |

• Q5_K_M：体积更小的日常量化，更适合 24GB 显存档位。
• Q8_0：更高保真量化，更接近 BF16 原权重。

来源关系

1. meta-models/Muse-Glimmer-30B — Meta Superintelligence Lab 发布的 Muse Glimmer 30B（Apache-2.0）。
2. gjtgjt/Muse-Glimmer-30B-heretic — 对该权重做 Heretic 消融后的版本。
3. 本仓库 — 上述 heretic 权重的 GGUF 量化，供 llama.cpp / LM Studio 等本地运行时使用。

Muse Glimmer 是约 30B 的多模态因果语言模型（输入文本 + 图像，输出文本），带感知编码器，面向本地智能体场景。

使用方法

需要支持 muse-glimmer 架构的 llama.cpp 构建。

# 命令行
llama-cli -m Muse-Glimmer-30B-heretic-Q5_K_M.gguf -ngl 99 -c 65536

# 服务
llama-server -m Muse-Glimmer-30B-heretic-Q5_K_M.gguf -ngl 99 -c 65536

LM Studio / Jan 等应用：直接指向本仓库，或把 .gguf 放进模型目录。

上游 Muse Glimmer 卡片推荐的采样参数：

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

推理强度可在系统提示中设置：Reasoning strength: low|medium|high|xhigh。编程和智能体任务建议使用 high 或 xhigh。

说明

• 这是消融后的模型，拒绝行为被有意降低。部署时请自行加护栏。
• 不面向 18 岁以下用户。
• 量化推理在边缘情况下可能与 BF16 heretic 权重略有差异。
• 架构、用途与限制见 Muse Glimmer 原卡片。

致谢

• 基础模型：Meta Superintelligence Lab — Muse Glimmer 30B
• 消融方法：p-e-w/heretic
• Heretic 权重：gjtgjt/Muse-Glimmer-30B-heretic

---

English

GGUF quantizations of gjtgjt/Muse-Glimmer-30B-heretic, a Heretic-abliterated (decensored) derivative of Meta's meta-models/Muse-Glimmer-30B.

Base model license: Apache-2.0.

Files

| File | Quant | Size |

|---|---|---|

| Muse-Glimmer-30B-heretic-Q5_K_M.gguf | Q5_K_M | 18.4 GiB |

| Muse-Glimmer-30B-heretic-Q8_0.gguf | Q8_0 | 27.6 GiB |

• Q5_K_M: smaller everyday quant, better fit for 24 GB class cards.
• Q8_0: higher-fidelity quant, closer to the BF16 source.

Lineage

1. meta-models/Muse-Glimmer-30B — Muse Glimmer 30B by Meta Superintelligence Lab (Apache-2.0).
2. gjtgjt/Muse-Glimmer-30B-heretic — Heretic abliteration of that checkpoint.
3. This repo — GGUF quantizations of the heretic weights for llama.cpp / LM Studio / similar runtimes.

Muse Glimmer is a ~30B multimodal causal LM (text + image in, text out) with a perception encoder, aimed at local agentic use.

Usage

Requires a llama.cpp build that supports the muse-glimmer architecture.

# CLI
llama-cli -m Muse-Glimmer-30B-heretic-Q5_K_M.gguf -ngl 99 -c 65536

# Server
llama-server -m Muse-Glimmer-30B-heretic-Q5_K_M.gguf -ngl 99 -c 65536

LM Studio / Jan / similar apps: point them at this repo or drop the .gguf into the models folder.

Upstream sampling defaults from the Muse Glimmer card:

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

Reasoning strength can be set in the system prompt as Reasoning strength: low|medium|high|xhigh. Use high or xhigh for coding and agentic tasks.

Notes

• This is an abliterated model: refusal behavior is reduced by design. Deploy with your own guardrails.
• Not intended for use by individuals under 18.
• Quantized inference can differ slightly from the BF16 heretic checkpoint in edge cases.
• See the base Muse Glimmer card for architecture, intended use, and limitations.

Credit

• Base: Meta Superintelligence Lab — Muse Glimmer 30B
• Abliteration method: p-e-w/heretic
• Heretic weights: gjtgjt/Muse-Glimmer-30B-heretic