Muse-Glimmer-30B-heretic-GGUF
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
- 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
- meta-models/Muse-Glimmer-30B — Meta Superintelligence Lab 发布的 Muse Glimmer 30B（Apache-2.0）。
- gjtgjt/Muse-Glimmer-30B-heretic — 对该权重做 Heretic 消融后的版本。
- 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。
1# 命令行
2llama-cli -m Muse-Glimmer-30B-heretic-Q5_K_M.gguf -ngl 99 -c 65536
3
4# 服务
5llama-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 = 1.0top_p = 0.95top_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
- Q5_K_M: smaller everyday quant, better fit for 24 GB class cards.
- Q8_0: higher-fidelity quant, closer to the BF16 source.
Lineage
- meta-models/Muse-Glimmer-30B — Muse Glimmer 30B by Meta Superintelligence Lab (Apache-2.0).
- gjtgjt/Muse-Glimmer-30B-heretic — Heretic abliteration of that checkpoint.
- 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.
1# CLI
2llama-cli -m Muse-Glimmer-30B-heretic-Q5_K_M.gguf -ngl 99 -c 65536
3
4# Server
5llama-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 = 1.0top_p = 0.95top_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
