OS-Software/Muse-Glimmer-30B-Uncensored-Heretic

VISIONPricing:Input $1.2 / Cached $0.04 / Output $4.4Concurrent Unit Cost:2Model Size:30BQuant:FP8Context Size:128kPublished:Sep 9, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

OS-Software/Muse-Glimmer-30B-Uncensored-Heretic is a 30-billion-parameter causal language model, a decensored version of Meta Superintelligence Lab's Muse-Glimmer-30B, created using Heretic v2.0.0. This model features a dedicated perception encoder for multimodal understanding and is optimized for autonomous agentic tasks on consumer hardware. Its primary differentiator is the substantial reduction of safety alignment, making it suitable for research into safety, alignment studies, and red-teaming.

Loading preview...

OS-Software/Muse-Glimmer-30B-Uncensored-Heretic Overview

This model is a 30-billion-parameter causal language model, derived from Meta Superintelligence Lab's Muse-Glimmer-30B, and processed with Heretic v2.0.0 to significantly reduce its safety alignment. It integrates a dedicated perception encoder for multimodal understanding, allowing it to accept interleaved text and images. Designed for autonomous agentic tasks, it runs efficiently on consumer hardware with a context length of 131,072+ tokens.

Key Differentiators & Capabilities

  • Decensored Nature: Substantially reduced safety alignment compared to the base model, making it prone to generating potentially harmful or inappropriate content. This is its core distinction.
  • Agentic Task Optimization: Excels in end-to-end agentic task completion, reliable tool use, multi-step reasoning, and failure recovery.
  • Multimodal Understanding: Processes both text and image inputs, enabling agents to interpret visual information like screenshots and charts.
  • Local Deployment Focus: Optimized for running on consumer hardware (24GB/32GB VRAM) using quantization techniques and speculative decoding for faster generation.
  • Performance: Outperforms Gemma4-31B and Qwen3.6-27B in several general agentic, agentic coding, and multimodal benchmarks, despite its decensored nature.

Intended Use Cases

  • Research and Experimentation: Primarily for safety research, alignment studies, and red-teaming of AI systems.
  • Local AI Agents: Developing multi-step planning, tool invocation, and long-horizon task execution on consumer devices.
  • Coding Agents: Tasks involving writing, debugging, and resolving software engineering problems.
  • Multimodal Reasoning: Applications requiring interpretation of visual and textual data for agentic environments.

Important Notice: Due to its reduced safety alignment, this model is intended for research and experimentation only. Users are solely responsible for evaluating its outputs and implementing safeguards.

Finetunes
2 models