DreamFast/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark

Model Overview

This model, Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark, is a HuggingFace safetensors recovery of a Qwen3.6-35B-A3B MoE hybrid model, originally distributed as a Q8_0 quantized GGUF by HauhauCS. It features a Qwen3_5MoeForConditionalGeneration architecture with 256 experts and 8 active per token. The recovery process involved converting the GGUF to safetensors with bit-exact verification and restoring Multi-Token Prediction (MTP) and vision encoder tensors from the official Qwen3.6-35B-A3B reference model.

Key Differentiators

• Abliterated Refusal Rate: Achieves a 0% refusal rate on harmful prompts, a significant reduction from the base model's 40%, indicating an "aggressive" and "uncensored" behavior. This was confirmed through vLLM testing.
• MoE Architecture: Leverages a Mixture-of-Experts (MoE) hybrid Gated DeltaNet + Gated Attention architecture for potentially efficient and high-quality generation.
• Provenance and Recovery: Meticulously recovered from a lossy Q8_0 quantization, with detailed tensor comparison against the base model revealing specific modifications to expert and projection weights responsible for the abliteration.

Use Cases

• Uncensored Content Generation: Ideal for applications where aggressive and unfiltered responses are desired, without the typical refusal mechanisms of base models.
• Research into Abliteration: Useful for studying the effects of model modifications on safety and refusal behaviors.
• High-Quality Generation with Specific Behavior: Provides 100% coherence on both harmful and benign prompts, matching the base model's generation quality while altering its refusal characteristics.