AMAImedia/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-heretic-NOESIS-BF16

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 1, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The AMAImedia/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-heretic-NOESIS-BF16 is a 27 billion parameter Qwen3.8-based model developed by AMAImedia, derived from gorbatjovy's refusal-removed version. This model, part of the NOESIS Professional Multilingual Dubbing Automation Platform, has undergone 'abliteration' to remove safety alignment and refusal behaviors. It retains the full vision-language tower and MTP speculative-decoding head, making it suitable for research into interpretability, refusal mechanisms, and red-teaming.

Loading preview...

Model Overview

This model, AMAImedia/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-heretic-NOESIS-BF16, is a 27 billion parameter Qwen3.8-based language model. It is a repack by AMAImedia, part of the NOESIS Professional Multilingual Dubbing Automation Platform, based on a version from gorbatjovy. The core differentiator is its "abliteration" via the Heretic method, which systematically removes safety alignment and refusal behaviors present in the base model.

Key Characteristics

  • Refusal-Removed: Utilizes Heretic's automated, KL-constrained abliteration to eliminate refusal mechanisms, making it comply with requests the base model would typically refuse.
  • Base Model: Built upon DavidAU/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1, which is a Cold-Fusion (GAIN + Unsloth) tune of Qwen/Qwen3.8-27B.
  • Architecture Preservation: Retains the full vision-language tower and the MTP speculative-decoding head, ensuring compatibility and functionality with the original base model's features.
  • Multilingual Support: Inherits multilingual capabilities, supporting English, Russian, Chinese, Japanese, Kazakh, and Vietnamese.
  • BF16 Format: Provided in BF16 (bfloat16) format, with W4A16 and Ninfer builds also available for smaller footprints.
  • Minimal Capability Drift: Abliteration is performed with a low KL divergence (0.0315), indicating that core capabilities are largely retained, and the vision tower remains byte-for-byte identical.

Intended Use Cases

This model is explicitly released for research purposes related to:

  • Interpretability: Studying how refusal mechanisms function within large language models.
  • Refusal-Mechanism Study: Analyzing the effects and methods of removing safety alignments.
  • Red-Teaming: Evaluating model robustness and identifying potential vulnerabilities.
  • Robustness Evaluation: Testing model behavior in scenarios where safety alignments are absent.

Disclaimer: Users are responsible for the model's output and must add their own safety and moderation layers if deploying to end-users, as its safety alignment has been removed.

Finetunes
1 models