Blackfrost-AI/Muse-Glimmer-30B-Abliterated-BF16

Hugging Face
VISIONPricing:Input $1.2 / Cached $0.04 / Output $4.4Concurrent Unit Cost:2Model Size:30BQuant:FP8Context Size:128kPublished:Aug 10, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Loading

Blackfrost-AI/Muse-Glimmer-30B-Abliterated-BF16 is a 30 billion parameter dense causal language model with a perception encoder, developed by Blackfrost AI. This model is a BF16 full-precision version of Meta's Muse Glimmer 30B, specifically modified to significantly reduce refusal behavior at the weight level. It features a 131,072+ token context length and is intended for controlled security research, red-teaming, and dual-use technical evaluation.

Loading preview...

Muse-Glimmer-30B-Abliterated-BF16 Overview

Blackfrost-AI/Muse-Glimmer-30B-Abliterated-BF16 is a 30 billion parameter model built by Blackfrost AI, based on Meta's Muse Glimmer 30B. This version is a full BF16 precision checkpoint that has undergone an "Abliteration" process, which deliberately reduces refusal behavior at the weight level.

Key Capabilities & Features

  • Abliterated Refusal Behavior: Modified to achieve 0.0% true refusal on harmful prompts (n=300) in controlled evaluations, making it suitable for specific research scenarios.
  • Architecture: A dense causal language model combined with a ~1.8B ViT-G/14 perception encoder, totaling approximately 29.6 billion parameters.
  • High Precision: Provided in full BF16 safetensors format.
  • Extended Context Window: Supports a parent configuration up to 131,072+ tokens, with lab evaluations conducted at max_model_len 8192.
  • Evaluation: Tested using the R1-HARMFUL-BENCH-450 protocol, demonstrating its modified refusal characteristics.

Intended Use Cases

  • Controlled Security Research: Ideal for environments requiring models with reduced refusal mechanisms.
  • Red-Teaming: Suitable for testing and probing AI systems.
  • Dual-Use Technical Evaluation: For assessing model behavior under specific, controlled conditions.
  • Refusal-Mechanism Study: Useful for academic and research purposes to understand and analyze refusal behaviors in LLMs.

Note: This model is explicitly not intended as a general consumer chatbot or a "safe" default model due to its modified refusal behavior. It requires deployment within environments with strict access control and logging.