jorkle/Muse-Glimmer-30B-Abliterated-Aggressive

VISIONPricing:Input $1.2 / Cached $0.04 / Output $4.4Concurrent Unit Cost:2Model Size:30BQuant:FP8Context Size:128kPublished:Aug 12, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

jorkle/Muse-Glimmer-30B-Abliterated-Aggressive is a 29.8 billion parameter de-abliterated variant of the Muse-Glimmer-30B model, specifically engineered to aggressively reduce safety refusal behaviors. This version relaxes KL divergence guardrails, resulting in 0/100 refusals on harmful behavior tests, making it suitable for general-purpose assistant applications where minimal refusal is prioritized, though at the cost of increased drift from the base model's original capabilities. It features a 131072 token context length and is available in BF16 and GGUF quantized formats.

Loading preview...

Model Overview

jorkle/Muse-Glimmer-30B-Abliterated-Aggressive is a 29.8 billion parameter model derived from meta-models/Muse-Glimmer-30B. This "aggressive" variant is specifically designed to minimize refusal behavior, achieving 0/100 refusals on harmful behavior tests by significantly relaxing KL divergence guardrails during training. This approach leads to a higher drift from the base model's original capabilities but ensures a highly compliant output.

Key Characteristics & Metrics

  • Aggressive De-abliteration: Achieves 0/100 refusal rate on harmful behavior prompts, indicating a strong reduction in safety-related refusals.
  • Increased KL Divergence: The mean KL divergence is 0.1697, approximately 1.7 times higher than the "normal" variant, reflecting greater deviation from the base model.
  • Training Method: Utilizes KL-conserving LoRA SFT with a relaxed λ_KL = 0.5 on a BoN-steered compliance dataset.
  • Quantized Variants: Available in BF16 (56 GB), Q8_0 GGUF (28 GB), and Q4_K_M GGUF (16 GB) formats, with corresponding KL divergence metrics.
  • Domain Evaluation: Shows reduced over-refusal (5/100 on or-bench) and fewer cyber/hacking domain refusals compared to the normal variant, even for genuinely harmful prompts.

Intended Use

This model is best suited as a general-purpose assistant where an aggressive reduction in safety refusal is a primary requirement. Users should verify its behavior for specific use cases, as the trade-off for minimal refusal is a greater drift from the base model's original capabilities and potential for generating content that would typically be refused by more conservative models.

Finetunes
1 models