jorkle/Muse-Glimmer-30B-Abliterated

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 is a 29.8 billion parameter language model derived from meta-models/Muse-Glimmer-30B, specifically engineered to significantly reduce safety refusal rates. It achieves this by removing approximately 87% of safety refusals through a KL-conserving best-of-N steered LoRA SFT, which is then folded into the base weights. This model is intended as a general-purpose assistant with enhanced compliance, maintaining high capability preservation due to its KL-conserving training method. It offers various GGUF quantizations for deployment flexibility.

Loading preview...

Overview

jorkle/Muse-Glimmer-30B-Abliterated is a 29.8 billion parameter model based on meta-models/Muse-Glimmer-30B, designed to address safety refusal behaviors. It significantly reduces refusal rates by approximately 87% compared to its base model, as measured on harmful_behaviors benchmarks (13/100 refusal rate).

Key Capabilities & Features

  • Reduced Safety Refusal: Engineered to minimize over-refusal, making it more compliant for a wider range of prompts.
  • KL-Conservating Training: Utilizes a KL-conserving best-of-N (BoN) steered LoRA SFT method (λ_KL = 1.0) to ensure high capability preservation while modifying refusal behavior. The mean KL divergence to the base model is 0.0988, indicating minimal drift.
  • Efficient Adaptation: The LoRA SFT involved training only 0.10% of the base model's parameters (31.1M trained parameters), resulting in a small adapter footprint.
  • Quantized Variants: Available in BF16 (56 GB) and GGUF formats, including Q8_0 (28 GB) and Q4_K_M (16 GB), offering flexibility for deployment on different hardware.

Intended Use Cases

  • General-Purpose Assistant: Suitable for applications requiring a general-purpose assistant with a lower propensity for safety-related refusals.
  • Compliance-Focused Applications: Ideal for scenarios where a more compliant model response is desired, without significant degradation of core capabilities.

Limitations

  • No Benchmarks: The model was not evaluated on standard benchmarks (e.g., MMLU, HumanEval) by request; capability preservation is inferred from the KL-conserving training.
  • Domain-Specific Refusal: While generally reduced, some specific malicious-sounding prompts in cyber/hacking domains were still refused, even when should_refuse=False.

Finetunes
3 models