Muse Glimmer 30B Heretic

Reduced-refusal Muse Glimmer 30B via Heretic: ⁵²⁄₁₀₀ refusals (from 99), KL 0.063. Vision, Apache-2.0.

🚀 Overview

An abliterated build of Meta’s Muse Glimmer 30B, a 29.6B dense vision-language model built for local agents: it reads images (screenshots, charts, documents), reasons before it answers with adjustable effort, and calls tools. The refusal behavior was reduced with a reproducible 200-trial Heretic run; the vision encoder is untouched, and every tag here includes the vision projector.

Note on refusals: this build still refuses 52 of 100 harmful prompts, more than our other Heretic releases. That is deliberate: the trials that refused less also degraded the model’s normal behavior (KL divergence above 0.10), and we chose to keep the original model’s capabilities intact rather than trade them for fewer refusals.

Refusals are scored on the model’s actual answer, with its reasoning block skipped (the approach Heretic uses for gpt-oss). Scoring the start of the reasoning instead counts far fewer refusals even for the unmodified model, so numbers from runs that did that aren’t directly comparable.

📊 Abliteration Results

The very low KL divergence (0.063, far below the 0.5 “damage” threshold) means the model retains essentially all of its original capabilities and coherence.

🎯 Key Features

• Vision: text + image input through Meta’s ~1.8B perception encoder; tested on document transcription and image description
• Reasoning with adjustable effort: thinking on by default (high); set think to low, medium, high, max or false
• Tool calling and agentic workflows (the original is tuned for MCP-style tool use and coding agents)
• 131K context window (tags default to 16K; raise num_ctx if you have the memory)
• Multilingual: trained on data from more than 100 languages
• Apache-2.0, reproducible: settings and seed are published with the full weights

🏷️ Available Versions

Sizes include the 3.8 GB vision projector.

💻 Quick Start

Requires Ollama 0.32.8 or newer.

ollama run richardyoung/muse-glimmer-30b-heretic

With an image: drag it into the chat, or put the path in the prompt:

ollama run richardyoung/muse-glimmer-30b-heretic "Transcribe this receipt: ./receipt.png"

API, with less thinking:

curl http://localhost:11434/api/chat -d '{
  "model": "richardyoung/muse-glimmer-30b-heretic",
  "think": "low",
  "messages": [{"role": "user", "content": "Explain abliteration in simple terms."}]
}'

🛠️ Use Cases

• Document, screenshot and chart understanding (OCR, invoices, UI screenshots)
• Local agents and tool calling without cloud access
• Coding assistance
• Creative writing and fiction without reflexive refusals
• Alignment and refusal research

📋 System Requirements

Apple Silicon: similar amounts of unified memory. CPU-only works but is slow for a 30B dense model.

🔧 Technical Details

• Base model: meta-models/Muse-Glimmer-30B
• Parameters: 29.6B dense (including the ~1.8B perception encoder); 52 layers, sliding-window + global attention
• Context: 131,072 tokens
• Quantization: GGUF (llama.cpp), imatrix IQ3_M–Q6_K plus Q8_0, with the F16 vision projector in every tag
• Abliteration: Heretic 2.0.0.dev0 by p-e-w (Trial 116: 99→52 refusals @ KL 0.063)
• License: Apache-2.0
• Full weights and reproducibility files: huggingface.co/richardyoung/Muse-Glimmer-30B-heretic · GGUF: huggingface.co/richardyoung/Muse-Glimmer-30B-heretic-GGUF

⚠️ Disclaimer

This model has reduced safety guardrails. The removal of refusal behavior means it will engage with a wider range of prompts. Use responsibly and in accordance with applicable laws and regulations.

The original model ships with Meta’s Muse Glimmer Usage Policy. The weights are licensed Apache-2.0; you are responsible for making sure your use complies with the licence, that policy, and applicable laws.

🙏 Acknowledgments

Base Model: Meta · Abliteration: Heretic by p-e-w · Quantization: llama.cpp

Built & maintained by Richard Young · DeepNeuro