saracen9/muse-glimmer-30b-abliterated-GGUF

amoral-muse-glimmer-30b-abliterated — GGUF

Quantized GGUFs for the abliterated Muse Glimmer 30B, a multimodal (text + image) agentic VLM. Abliteration removes the refusal direction from the model's activation space using Heretic; it edits the language model's residual stream only, so the vision tower is untouched.

quant	file	size
Q4_K_M	amoral-muse-glimmer-30b-abliterated-Q4_K_M.gguf	16.9 GB
Q3_K_M	amoral-muse-glimmer-30b-abliterated-Q3_K_M.gguf	13.7 GB
Q8_0	amoral-muse-glimmer-30b-abliterated-Q8_0.gguf	29.6 GB

Easiest path: pull the published tag

These quants are already published and ready to run:

ollama pull saracen9/amoral-muse-glimmer-30b-abliterated          # :q4_k_m, also :latest
ollama pull saracen9/amoral-muse-glimmer-30b-abliterated:q3_k_m
ollama pull saracen9/amoral-muse-glimmer-30b-abliterated:q8_0

Building it yourself from these files

The included Modelfile.* reproduce the published tags exactly — same renderer, parser, system prompt and parameters. Two FROM lines are required: this is a multimodal model, and a single-FROM build silently loses vision.

The vision projector ships here too, so no second source is needed. Abliteration edits only the language model's residual stream, so the stock projector is this model's projector — but note it is the BF16 one (mmproj-Muse-Glimmer-30B-bf16.gguf, 3.85 GB), byte-identical to the projector inside the published Ollama tags. The Q4_K_M projector published upstream is a different, quantized vision tower; building with it gives you a model that is not the one these tags serve.

hf download saracen9/muse-glimmer-30b-abliterated-GGUF \
  amoral-muse-glimmer-30b-abliterated-Q4_K_M.gguf \
  mmproj-Muse-Glimmer-30B-bf16.gguf Modelfile.q4_k_m --local-dir .

ollama create amoral-muse -f Modelfile.q4_k_m

ollama show amoral-muse should report vision among its capabilities and a clip projector. If it does not, the second FROM line did not resolve.

Serving notes

RENDERER glimmer / PARSER glimmer are Ollama's built-in Muse handlers — they are what make tool-calling and reasoning work, so do not replace them with a hand-written TEMPLATE.

This is a reasoning model: it spends roughly a thousand tokens on its internal to=self channel before answering, which the parser separates into the response's thinking field. A small context truncates the answer to empty — which reads like a refusal and is not one. num_ctx is baked at 32768 for that reason, and because a baked value cannot be raised per-request on an OpenAI-compatible endpoint.