AMAImedia/Qwen3.8-27B-Uncensored-Aggressive-NOESIS-BF16

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 31, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

AMAImedia/Qwen3.8-27B-Uncensored-Aggressive-NOESIS-BF16 is a 27 billion parameter, multilingual, multimodal (image-text-to-text) model based on the Qwen3.8 architecture. Developed by AMAImedia as part of the NOESIS platform, this model is a de-refused build of philbert440's Qwen3.8-27B-Uncensored-Aggressive, optimized for increased openness and improved reasoning. It retains the vision tower and MTP speculative-decoding head, making it suitable for tasks requiring multimodal understanding and less restrictive instruction following.

Loading preview...

Model Overview

AMAImedia/Qwen3.8-27B-Uncensored-Aggressive-NOESIS-BF16 is a 27 billion parameter, multilingual, multimodal language model developed by AMAImedia as part of the NOESIS Professional Multilingual Dubbing Automation Platform. This model is a refined version of philbert440's Qwen3.8-27B-Uncensored-Aggressive, specifically engineered to be "de-refused" or "abliterated" for greater openness in instruction following.

Key Characteristics

  • Multimodal Capabilities: Preserves the vision tower and MTP speculative-decoding head, enabling image-text-to-text processing.
  • Enhanced Openness: Utilizes a specific refusal-vector orthogonalization method (alpha=1.15) to achieve peak openness, improving instruction adherence and reasoning compared to previous aggressive builds.
  • Multilingual Support: Supports a wide array of languages, including English, Russian, Chinese, Vietnamese, Japanese, and many others.
  • BF16 Precision: Trained and evaluated using BF16 precision.

Evaluation Highlights

Evaluations show this model achieves a high openness score of 0.88, while maintaining strong factual accuracy (1.00) and GSM8K performance (0.85). It demonstrates improved openness and reasoning compared to earlier aggressive builds, with reduced confabulation.

Deployment Options

Various quantization formats are available for deployment, including:

  • -W4A16-AWQ: 4-bit weight AWQ with MTP head.
  • -NVFP4: NVFP4 weight/activation with MTP head.
  • -GGUF: llama.cpp GGUF quants, including vision mmproj and MTP head.

Use Case Considerations

This model is designed for applications requiring a highly open, multimodal, and multilingual LLM that will follow instructions it might otherwise decline. Users should exercise responsibility and ensure compliance with applicable laws due to its uncensored nature.

Finetunes
1 models