Architecture: Qwen 3.5 35B Mixture-of-Experts Transformer with MTP Heads
Merging Technique: MoE Weight Patching & Tensor Grafting
Constituent Models:

• Model A: [object Object]

Methodology: MTP prediction layers stripped of layout corruption and re-grafted cleanly for local engine compatibility.

> Key Upgrades Over Base Model:

• Uncensored Freedom: Alignment layers ablated to prevent refusal walls on complex tasks.
• MTP Speculative Decoding: Multi-token prediction heads enable ultra-fast inference on supported backends.
• MoE Efficiency: Sparse activation ensures high intelligence with reasonable VRAM consumption.

// Note: Arena Elo and head-to-head winrates updated continuously as evaluation telemetry processes.

> Recommended Settings

• Temperature: 0.2 - 0.7
• Top-P: 0.95
• Backend Engines: Compatible with llama.cpp, vLLM, Ollama, LM Studio, KoboldCPP

• [+] llama.cpp: Native support across all quantizations.
• [+] Ollama / LM Studio: Full GGUF compatibility.

Disclaimer: Qwen3.5-35B-A3B-Uncensored-HauhauCS-Aggressive-MTP is provided for research and sovereign local deployment. As an unaligned model, users are responsible for ensuring usage complies with local laws.

Credits: Gratitude to original base model authors (HauhauCS/Qwen3.5-35B-A3B-Uncensored-HauhauCS-Aggressive) and open-source AI community tools.