TrevorJS/Muse-Glimmer-30B-uncensored

VISIONPricing:Input $1.2 / Cached $0.04 / Output $4.4Concurrent Unit Cost:2Model Size:30BQuant:FP8Context Size:128kPublished:Aug 11, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

TrevorJS/Muse-Glimmer-30B-uncensored is a 30 billion parameter uncensored version of the Muse-Glimmer model, developed by TrevorJS. This model has been specifically modified to remove refusal behavior, achieving a significant reduction in harmful refusals from 85.3% to 2.0% while maintaining agentic validation. It is a dense multimodal model with 52 text decoder layers and a 1.9B ViT-G/14 vision tower, optimized for applications requiring direct and unconstrained responses.

Loading preview...

Muse-Glimmer-30B-uncensored Overview

TrevorJS/Muse-Glimmer-30B-uncensored is a 30 billion parameter language model derived from the meta-models/Muse-Glimmer-30B base model. Its primary distinction is the removal of refusal behavior, achieved through a norm-preserving biprojected abliteration method. This process specifically targets and eliminates the 'refusal direction' from the model's residual-write matrices, ensuring that the model provides direct answers without generating refusal statements or deflections.

Key Capabilities & Differentiators

  • Significantly Reduced Refusals: Demonstrates a reduction in harmful refusals from 85.3% to 2.0% and eliminates over-refusal on harmless prompts.
  • Preserved Agentic Behavior: Maintains prompt-injection resistance and scope adherence, with a slight improvement in irreversible-action confirmation (from 7/30 to 8/30 probes).
  • Multimodal Architecture: Based on a dense multimodal model featuring 52 text decoder layers and a separate 1.9B ViT-G/14 vision tower, though only the text pathway was modified for uncensoring.
  • Norm-Preserving Abliteration: Utilizes a sophisticated method to remove refusal tendencies while preserving the original norm of the weight matrices, minimizing degradation of other capabilities.

Ideal Use Cases

  • Applications requiring direct responses: Suitable for scenarios where the model must provide information or complete tasks without generating safety-related refusals.
  • Research into model safety and alignment: Provides a valuable baseline for studying the impact of refusal mechanisms and methods for their removal.
  • Creative and unconstrained content generation: Useful for tasks that benefit from a model that does not self-censor or deflect based on perceived harmfulness.