darkc0de/Muse-Glimmer-30B-heretic
darkc0de/Muse-Glimmer-30B-heretic is a 30-billion-parameter decensored version of Meta Superintelligence Lab's Muse-Glimmer-30B, created using Heretic v1.4.0. This multimodal causal language model features a dedicated perception encoder and a 131,072+ token context length, purpose-built for autonomous agentic tasks on consumer hardware. It integrates multi-step reasoning, reliable tool use, multimodal understanding, and failure recovery, excelling in local AI agents, coding agents, and multimodal reasoning. The model demonstrates significantly reduced refusals (11/100) compared to its original counterpart (59/100) while maintaining strong performance in agentic benchmarks.
Loading preview...
Muse-Glimmer-30B-heretic: Decensored Agentic AI for Local Deployment
darkc0de/Muse-Glimmer-30B-heretic is a 30-billion-parameter variant of Meta Superintelligence Lab's Muse-Glimmer-30B, specifically modified using Heretic v1.4.0 to be a decensored model. This version significantly reduces refusals, showing 11/100 refusals compared to the original's 59/100, while maintaining the core capabilities of the base model. It is a multimodal causal language model with a dedicated perception encoder, designed for autonomous agentic tasks on consumer hardware.
Key Capabilities & Features
- Decensored Output: Achieves a much lower refusal rate, offering broader utility for various applications.
- Autonomous Agentic Tasks: Excels in end-to-end task completion, reliable tool use, multi-step reasoning, and failure recovery.
- Multimodal Understanding: Processes interleaved text and images via a dedicated perception encoder, enabling agents to interpret visual data like screenshots and charts.
- Optimized for Local Deployment: Engineered to run efficiently on consumer hardware (24GB/32GB VRAM) using 4-bit quantization with minimal performance degradation (0.2-1.0%).
- Faster Generation: Incorporates speculative decoding with a DFlash drafter model, achieving up to 3.1x speedup on RTX 5090 and 1.5-1.8x on Apple M-series chips.
- Broad Agentic Compatibility: Works with OpenClaw, Hermes Agent, and other orchestration patterns.
- Multilingual Support: Trained on data from over 100 languages.
Performance Highlights
The model demonstrates strong performance across various agentic benchmarks, often outperforming Gemma4-31B and Qwen3.6-27B in its size class. Notable scores include 75.5 on MCP Atlas, 74.6 on DeepSearch QA, and 51.2 on SWE-Bench Pro. It also shows competitive results in multimodal benchmarks like Charxiv Reasoning (78.8).
Ideal Use Cases
- Local AI Agents: For multi-step planning, tool invocation, and long-horizon task execution on consumer devices.
- Coding Agents: Writing, debugging, and resolving software engineering tasks.
- Multimodal Reasoning: Interpreting visual information alongside text for agentic and information-rich environments.
- Synthetic Data Generation: Creating high-quality training data.
- LLM-as-a-Judge: Evaluating outputs from other models.