This project provides an "uncensored" and capability-enhanced version of the Qwen3.6-27B language model, specifically optimized for NVIDIA DGX Spark and Blackwell hardware using NVFP4 quantization. It targets researchers and power users seeking to bypass safety alignment restrictions and leverage the model's full potential, offering significant performance improvements and restored reasoning capabilities.

How It Works

The core of the project is a multi-stage "lossless abliteration" process. It begins with repairing an upstream defect in Qwen3.6's GatedDeltaNet (Mamba-style) layers. Subsequently, it employs abliterix v1.4, a multi-objective optimization tool, to surgically remove the model's refusal direction while minimizing capability degradation, validated against a 10-axis spot-check to prevent word-salad outputs. Finally, it applies NVFP4 quantization, a 4-bit floating-point format designed for Blackwell-era hardware, preserving BF16 accuracy with true 4-bit compute throughput. Critical components like SSM layers retain BF16 precision to ensure correctness.

Quick Start & Requirements

• DGX Spark (Recommended): Utilizes a vllm-aeon-ultimate-dflash container (ghcr.io/aeon-7/vllm-aeon-ultimate-dflash:qwen36-v4) with docker-compose.spark-xs.yml. Requires Docker, NVIDIA GPUs (DGX Spark/GB10), and CUDA. Setup involves hf auth login, hf download for model weights, and docker compose up -d. First boot takes ~10-12 minutes due to caching.
• A100/H100: Uses BF16 weights with vanilla vLLM via docker-compose.bf16.yml. Requires Docker, NVIDIA GPUs (A100/H100), and CUDA.
• Other Hardware: Recipes and tuned configurations are available in the other-hardware/ directory.
• Links:
  • BF16 Model: https://huggingface.co/AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-BF16
  • NVFP4 Model: https://huggingface.co/AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-NVFP4
  • DFlash Container: https://github.com/AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-DFlash/pkgs/container/vllm-aeon-ultimate-dflash

Highlighted Details

• Performance: Achieves an average +258% single-stream decode speed increase on DGX Spark v4 DFlash compared to the raw baseline, reaching up to 37.56 tok/s.
• Capability Preservation: Maintains a KL divergence of 0.000492 from the base model, significantly below the typical "capability damage threshold," with 10/10 coherent results across math, code, reasoning, and knowledge tasks.
• Uncensored: Demonstrates 0 refusals out of 100 tested harmful prompts.
• NVFP4 Quantization: Offers BF16-level accuracy with true 4-bit compute throughput on Blackwell hardware, preserving critical SSM layers in BF16.

Maintenance & Community

The project leverages and collaborates with upstream maintainers of abliteration tooling and vLLM. Specific community links like Discord or Slack are not explicitly provided in the README.

Licensing & Compatibility

The model is released under the Apache 2.0 license, inherited from the base Qwen3.6-27B model. This license generally permits commercial use and integration into closed-source projects.

Limitations & Caveats

This is an uncensored model; users are solely responsible for implementing downstream safety layers (input validation, output filtering, moderation) as the model will execute instructions without internal judgment. NVFP4 quantization offers no throughput advantage on pre-Blackwell hardware (e.g., A100/H100) where it dequantizes to BF16. The DDTree v5 research track is experimental and not production-ready.