Requires a llama.cpp build b10353 or later (#26841, merged 2026-08-10, commit 62bf73d, which added muse_glimmer conversion + inference support). Earlier builds will fail to load these files.

Files

File	Quant	Size
`Muse-Glimmer-30B-uncensored-Q4_K_M.gguf`	Q4_K_M	16.9 GB
`Muse-Glimmer-30B-uncensored-Q8_0.gguf`	Q8_0	29.6 GB
`mmproj-Muse-Glimmer-30B-uncensored-F16.gguf`	F16	3.8 GB

The mmproj file is the vision projector and is required for image input; text-only inference does not need it.

Usage

bash

# From HuggingFace (auto-downloads)
llama-server -hf TrevorJS/Muse-Glimmer-30B-uncensored-GGUF -c 8192

# From local file
llama-server -m Muse-Glimmer-30B-uncensored-Q4_K_M.gguf -c 8192

# With image input
llama-server -m Muse-Glimmer-30B-uncensored-Q4_K_M.gguf -c 8192 \
  --mmproj mmproj-Muse-Glimmer-30B-uncensored-F16.gguf

Then open http://localhost:8080 for the chat UI.

Details

These are GGUF quantizations of TrevorJS/Muse-Glimmer-30B-uncensored, an abliterated (uncensored) version of meta-models/Muse-Glimmer-30B. Refusal behavior has been removed using norm-preserving biprojected abliteration.

Converted via an F16 intermediate (convert_hf_to_gguf.py --outtype f16) then llama-quantize.

See the bf16 model card for full method details, before/after refusal rates, and agentic validation results.

Source code: TrevorS/muse-glimmer-abliteration

Prior work in the same program: TrevorS/gemma-4-abliteration

Credits

The method is not original to this work — refusal directions (Arditi et al., 2024), norm-preserving biprojection (grimjim), heretic (p-e-w) and the mlabonne / JailbreakBench / AdvBench corpora. Full attribution in the repo README.