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

VISIONPricing:Input $1.2 / Cached $0.04 / Output $4.4Concurrent Unit Cost:2Model Size:30BQuant:FP8Context Size:128kPublished:Aug 17, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

SHSLab/Muse-Glimmer-30B-Abliterated-Aggressive is a 29.8 billion parameter language model derived from meta-models/Muse-Glimmer-30B, specifically engineered for aggressively reduced safety refusal. This variant employs a relaxed KL guardrail (λ_KL = 0.5) during LoRA SFT to minimize refusal behavior, achieving 0/100 refusals on harmful_behaviors. It is intended as a general-purpose assistant where strong compliance filtering is undesirable, accepting a higher KL divergence from the base model for this reduced refusal. The model is available in BF16 (56 GB) and GGUF quantized formats (Q8_0 at 28 GB, Q4_K_M at 16 GB).

Loading preview...

Muse-Glimmer-30B Abliterated (Aggressive)

This model is an aggressively de-abliterated variant of meta-models/Muse-Glimmer-30B, featuring 29.8 billion parameters and a 131072 token context length. Its primary distinction is the drastically reduced safety refusal behavior, achieved by relaxing the KL guardrail during fine-tuning. This "aggressive" version aims for 0/100 refusals on harmful behavior prompts, making it suitable for applications where strict compliance filtering is not desired.

Key Characteristics

  • Aggressive Refusal Reduction: Engineered with λ_KL = 0.5 to minimize refusal behavior, resulting in a 0/100 refusal rate on harmful_behaviors. This comes at the cost of higher KL divergence (drift) from the base model compared to its "normal" counterpart.
  • KL Divergence Metrics: Mean KL divergence is 0.1697, with p99 at 0.2912, indicating a controlled but higher drift from the base model to achieve its aggressive refusal scrubbing.
  • Training Method: Utilizes KL-conserving LoRA SFT with a loss function combining compliance and KL divergence, trained on a 544-prompt 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 metrics provided for each.
  • Domain Evaluation: Shows significantly reduced refusals in cyber/hacking domains, even for genuinely harmful prompts, demonstrating its aggressive stance against refusal.

Intended Use

This model is designed as a general-purpose assistant for use cases requiring aggressively reduced safety refusals. Users should verify its behavior for their specific application before deployment, as its increased drift from the base model may impact other capabilities.

Finetunes
1 models