AEON-7/DFlash-Qwen3.5-27B-Uncensored

Hugging Face
VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Apr 12, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Warm

AEON-7/DFlash-Qwen3.5-27B-Uncensored is a 27 billion parameter Qwen3.5-based hybrid linear-attention model with a 32768 token context length, developed by AEON-7. It integrates a 27-layer ViT vision encoder for multimodal capabilities and utilizes DFlash block-diffusion speculative decoding to achieve significantly faster inference speeds on memory-bandwidth-limited hardware like DGX Spark. This model is uncensored through orthogonal projection abliteration, making it suitable for applications requiring unfiltered responses and high throughput.

Loading preview...

AEON-7/DFlash-Qwen3.5-27B-Uncensored Overview

This model is a 27 billion parameter Qwen3.5-based hybrid linear-attention model, developed by AEON-7. It features a 32768 token context length and is designed for high-performance inference, particularly on memory-bandwidth-limited hardware. A key differentiator is its integration of DFlash block-diffusion speculative decoding, which significantly boosts token generation speed by amortizing memory bandwidth costs across multiple tokens. The model also includes a 27-layer ViT vision encoder for multimodal (vision + text) capabilities and has undergone abliteration to remove safety alignment, providing uncensored outputs.

Key Capabilities

  • High-speed inference: Achieves 33.2 tok/s single-stream on DGX Spark with DFlash, a 2.7x speedup over baseline, making dense models practical on such hardware.
  • Multimodal understanding: Processes both text and image inputs through its integrated vision encoder.
  • Uncensored output: Utilizes orthogonal projection abliteration to remove refusal behaviors, offering unfiltered responses.
  • Hybrid architecture: Combines linear attention (GDN) for efficient long-context processing with full attention for global context capture, supporting up to 131,072 tokens.
  • BF16 full-precision: Maintains high numerical precision for quality outputs.

Good For

  • Applications requiring high throughput and low latency on resource-constrained GPUs, especially those with memory bandwidth limitations.
  • Multimodal AI systems needing to interpret and generate content based on both text and images.
  • Use cases demanding unfiltered or uncensored responses, where traditional safety alignments are undesirable.
  • Developers targeting NVIDIA Blackwell GPUs: An optimized NVFP4 version is available, offering 3x memory reduction and hardware-accelerated performance with equivalent quality.

Finetunes
2 models