0xA50C1A1/Muse-Glimmer-30B-abliterix

VISIONPricing:Input $1.2 / Cached $0.04 / Output $4.4Concurrent Unit Cost:2Model Size:30BQuant:FP8Context Size:128kPublished:Sep 13, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

The 0xA50C1A1/Muse-Glimmer-30B-abliterix is a 30 billion parameter language model, derived from the meta-models/Muse-Glimmer-30B architecture. This version has been decensored using Abliterix v1.12.2, specifically engineered to reduce refusal rates while maintaining a low KL divergence from the original model. It is optimized for applications requiring less restrictive content generation, offering a significantly lower propensity for refusals compared to its base model.

Loading preview...

Overview

This model, 0xA50C1A1/Muse-Glimmer-30B-abliterix, is a 30 billion parameter language model based on the meta-models/Muse-Glimmer-30B architecture. Its primary distinction is that it has been decensored using the Abliterix v1.12.2 tool. This modification aims to reduce the model's tendency to refuse certain prompts or content generation requests.

Key Characteristics

  • Decensored Version: Specifically engineered to lower refusal rates compared to the original Muse-Glimmer-30B model.
  • Low KL Divergence: Achieves a KL divergence of 0.0094, indicating that its output distribution remains very close to the original model despite the decensoring.
  • Reduced Refusals: Demonstrates a significant reduction in refusals, with only 10 refusals out of 100 prompts, compared to 98 refusals out of 100 for the original model.
  • Steering Parameters: Utilizes specific steering parameters for attn.o_proj and mlp.down_proj to achieve its decensored behavior.

Good For

  • Use cases requiring a language model with fewer content restrictions or a lower propensity for refusing prompts.
  • Applications where maintaining the core capabilities of the Muse-Glimmer-30B model is important, but with enhanced flexibility in content generation.
  • Developers looking for a 30B parameter model that offers a more permissive output without drastically altering the original model's statistical properties.

Finetunes
2 models