MuXodious/Muse-Glimmer-30B-SOMPOA-heresy
Muse-Glimmer-30B-SOMPOA-heresy by MuXodious is a 30-billion parameter fine-tuned Muse-Glimmer-30B model, developed using P-E-W's Heretic abliteration engine with Self-Organizing Maps & Magnitude-Preserving Orthogonal Ablation. This model is specifically engineered to reduce policy alignment and enable uncensored generations, offering a 131,072+ token context length. It is optimized for autonomous agentic tasks, multi-step reasoning, and reliable tool use on consumer hardware, while maintaining strong performance on agentic and multimodal benchmarks.
Loading preview...
Muse-Glimmer-30B-SOMPOA-heresy: Ablated for Agentic Autonomy
This model is a 30-billion parameter fine-tune of the Muse-Glimmer-30B base model, created by MuXodious using P-E-W's Heretic abliteration engine. Its primary distinction lies in the application of Self-Organizing Maps & Magnitude-Preserving Orthogonal Ablation (SOMPOA) to significantly reduce inherent policy alignment, aiming for more uncensored and flexible generations. While some residual policy alignment remains, the model provides specific jailbreak prompts and chat templates to facilitate fully uncensored outputs.
Key Capabilities & Features
- Reduced Policy Alignment: Engineered to minimize built-in content moderation, allowing for broader generative freedom.
- Agentic Task Optimization: Designed for autonomous agentic tasks, including multi-step reasoning, reliable tool use, and failure recovery.
- Multimodal Understanding: Features a dedicated perception encoder for interleaved text and image inputs, enabling agents to interpret visual data.
- Local Deployment Focus: Optimized for efficient operation on consumer hardware, utilizing 4-bit quantization to fit within 24-32 GB VRAM.
- Accelerated Generation: Incorporates speculative decoding with a DFlash drafter model, achieving up to 3.1x speedup on an Nvidia RTX 5090.
- Strong Benchmark Performance: Outperforms Gemma4-31B and Qwen3.6-27B in several agentic, coding, and multimodal benchmarks like MCP Atlas, DeepSearch QA, and SWE-Bench Pro.
- High Context Length: Supports a context length of 131,072+ tokens.
Good For
- Local AI Agents: Developing and deploying AI agents for complex, multi-step tasks on consumer devices.
- Coding Agents: Automating software engineering tasks, including code writing, debugging, and problem resolution.
- Uncensored Content Generation: Use cases requiring outputs free from typical LLM content restrictions, with provided tools for mitigation.
- Multimodal Applications: Agents that need to process and reason over both text and image inputs.
- Synthetic Data Generation: Creating high-quality training data for various downstream model developments.