richardyoung/Muse-Glimmer-30B-heretic

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

The richardyoung/Muse-Glimmer-30B-heretic is a 30-billion-parameter multimodal (text + image) vision-language model, based on Meta's Muse-Glimmer-30B, with a 131,072+ token context length. It has been decensored using the Heretic v2.0.0.dev0 method, significantly reducing refusals while retaining original capabilities. This model is optimized for autonomous agentic tasks, including multi-step reasoning, reliable tool use, and multimodal understanding, designed for local deployment on consumer hardware.

Loading preview...

Overview

This model, richardyoung/Muse-Glimmer-30B-heretic, is a decensored version of Meta's 29.6B parameter Muse-Glimmer-30B, a dense vision-language model (text + image input) with a 131,072+ token context length. It was created using the Heretic v2.0.0.dev0 method, which surgically modifies the model's weights to reduce refusals without retraining. The process involved 200 trials to find the optimal balance, resulting in a 47% reduction in refusals (from 99/100 to 52/100 harmful prompts) with a very low KL divergence (0.063), indicating minimal degradation of original capabilities.

Key Capabilities

  • Reduced Refusals: Significantly fewer refusals compared to the original Muse-Glimmer-30B, enabling more unrestricted local inference.
  • Multimodal Understanding: Accepts interleaved text and image inputs, allowing agents to interpret screenshots, charts, and documents.
  • Agentic Task Completion: Excels at end-to-end agentic tasks, including multi-step planning, sequential tool invocation, and failure recovery.
  • Reliable Tool Use: Handles a wide range of function calls with precise schemas throughout extended workflows.
  • Optimized for Local Deployment: Designed to run efficiently on consumer hardware (e.g., 24GB or 32GB VRAM) using quantization techniques and speculative decoding for faster generation.
  • Controllable Reasoning: Supports adjustable reasoning strengths (low, medium, high, xhigh) via system prompts.

Good For

  • Research and Education: Ideal for studying alignment techniques and exploring unrestricted local inference.
  • Local AI Agents: Developing and deploying autonomous agents for complex tasks on consumer devices.
  • Coding Agents: Tasks involving writing, debugging, and resolving software engineering problems (e.g., SWE-Bench style workflows).
  • Multimodal Applications: Scenarios requiring interpretation of both text and visual information, such as document analysis or interactive interfaces.
  • Synthetic Data Generation: Creating high-quality training data for downstream model development.
  • LLM-as-a-Judge Evaluation: Utilizing the model to evaluate outputs from other language models.

Finetunes
2 models