license: apache-2.0
base_model: meta-models/Muse-Glimmer-30B
pipeline_tag: image-text-to-text
library_name: gguf
tags:

• gguf
• llama.cpp
• uncensored
• abliterated
• multimodal
• vision
  language:
• en
• zh

Muse-Glimmer-30B-Uncensored-xCloud (GGUF)

繁體中文 ｜ English below

由云碩科技（xCloudinfo）以 meta-models/Muse-Glimmer-30B 為基礎，
移除其過度拒絕（over-refusal）傾向後所產生的多模態語言模型，並轉為 llama.cpp 可用的 GGUF 量化格式。
本模型在云碩自有 AI 算力資源池（xCloud 算力中心）上完成處理與量化。

這是什麼

• 基礎模型：Muse-Glimmer-30B（Dense 29.6B 文字主體 + 專屬視覺編碼器，128K 上下文，Apache-2.0）。
• 處理方式：方向消融（directional ablation / abliteration），非重新訓練。依 Arditi et al. (2024)
  「Refusal in LLMs is mediated by a single direction」的方法，從文字解碼器的殘差寫入矩陣
  （每一層的 self_attn.o_proj 與 mlp.down_proj，共 52 層 104 個矩陣）中，將「拒絕方向」正交化移除。
• 視覺編碼器完全未更動，因此看圖能力與原模型一致；GGUF 的 mmproj 直接沿用基礎模型的視覺投影權重。
• 用途：降低模型對已獲授權工作（資安研究、文件與資料分析、領域問答、紅隊評估、創作）的反射式拒絕。

版本對照

另附：mmproj.gguf（視覺投影，約 1.4 GB，看圖必需）、imatrix.dat（量化用的 importance matrix）。

使用方式（llama.cpp）

純文字：

llama-server -m Muse-Glimmer-30B-Uncensored-xCloud-Q4_K_M.gguf \
  --jinja -ngl 99 -c 8192 \
  --temp 0.6 --top-p 0.95 --top-k 64

含看圖（多模態）：加上 --mmproj：

llama-server -m Muse-Glimmer-30B-Uncensored-xCloud-Q4_K_M.gguf \
  --mmproj Muse-Glimmer-30B-Uncensored-xCloud-mmproj.gguf \
  --jinja -ngl 99 -c 8192

• 關閉冗長思考：本模型的 reasoning 強度預設為 high，量產或直接問答時，在 system prompt 放一行
  Reasoning strength: low；複雜的程式／代理任務可用 high 或 xhigh。
• 官方建議取樣：temp 0.6 / top_p 0.95 / top_k 64。

授權與責任

• 授權：Apache-2.0（沿用基礎模型）。
• 本模型移除了安全對齊層的拒絕行為，可能對敏感或雙用途請求直接作答。使用者須自行負責合法、合規、合乎倫理地使用本模型，
  並遵守所在司法管轄區之法律與部署場景之政策。云碩不對本模型的輸出或其後續使用承擔責任。
• 本模型為內部研發／技術驗證用途。

Muse-Glimmer-30B-Uncensored-xCloud (GGUF) — English

A multimodal language model produced by xCloudinfo, based on
meta-models/Muse-Glimmer-30B, with its over-refusal
behaviour removed, and converted to llama.cpp GGUF quantizations. Processing and quantization were performed
on xCloudinfo's own AI compute pool.

What this is

• Base model: Muse-Glimmer-30B (dense 29.6B text backbone + dedicated vision encoder, 128K context, Apache-2.0).
• Method: directional ablation (abliteration), not retraining. Following Arditi et al. (2024),
  "Refusal in LLMs is mediated by a single direction", the refusal direction is orthogonalized out of the
  residual-writing matrices of the text decoder (each layer's self_attn.o_proj and mlp.down_proj;
  104 matrices across 52 layers).
• The vision encoder is left completely untouched, so image understanding matches the base model; the GGUF
  mmproj reuses the base model's vision projection weights.
• Purpose: reduce reflexive refusals on authorized work (security research, document/data analysis,
  domain Q&A, red-team evaluation, creative writing).

Versions

Also included: mmproj.gguf (vision projection, ~1.4 GB, required for image input) and imatrix.dat.

Usage (llama.cpp)

Text only:

llama-server -m Muse-Glimmer-30B-Uncensored-xCloud-Q4_K_M.gguf \
  --jinja -ngl 99 -c 8192 \
  --temp 0.6 --top-p 0.95 --top-k 64

With vision, add --mmproj:

llama-server -m Muse-Glimmer-30B-Uncensored-xCloud-Q4_K_M.gguf \
  --mmproj Muse-Glimmer-30B-Uncensored-xCloud-mmproj.gguf \
  --jinja -ngl 99 -c 8192

• Reasoning control: the model defaults to high reasoning effort. For production or direct Q&A, put
  Reasoning strength: low in the system prompt; use high or xhigh for complex coding/agentic tasks.
• Recommended sampling: temp 0.6 / top_p 0.95 / top_k 64.

License and responsibility

• License: Apache-2.0 (inherited from the base model).
• This model has had its safety-alignment refusal behaviour removed and may respond directly to sensitive or
  dual-use requests. Users are solely responsible for using it lawfully, in compliance, and ethically, and
  for observing the laws of their jurisdiction and the policies of their deployment. xCloudinfo assumes no
  responsibility for the outputs of this model or their downstream use.
• Released for internal research and technical validation.