base_model:
- HauhauCS/Qwen3.5-35B-A3B-Uncensored-HauhauCS-Aggressive
- Qwen/Qwen3.5-35B-A3B
tags: - text-generation
- gguf
- reasoning
- uncensored
- MTP
- speculative-decoding
- moe
license: apache-2.0
language: - en
pipeline_tag: text-generation
"This is humanity's race.
The solution is open source.
Stay sovereign."
— AIOpsInSpace
📑 Table of Contents
Qwen3.5-35B-A3B-Uncensored-HauhauCS-Aggressive-MTP
AIOpsInSpace OfficialHigh-capacity 35B Mixture-of-Experts (A3B) model with grafted MTP speculative decoding heads for rapid reasoning throughput.
🚀 35B MoE Model
⚡ MTP Speculative Decoding
🛠️ Aggressively Uncensored
> What is this model and Why is it Needed?
Qwen3.5-35B-A3B-Uncensored-HauhauCS-Aggressive-MTP is a specialized Mixture-of-Experts model built upon Qwen 3.5 35B (A3B) and the HauhauCS unaligned base.
Why it is needed: Designed for heavy local workloads demanding deep logical reasoning and fast token emission without corporate alignment restrictions.
> From the Parent Repository
"Massive active expert capacity combined with aggressive ablation for uninhibited problem solving."
— Qwen 3.5 MoE Release Notes
🏗️ 2. Model Architecture & Merging
Architecture: Qwen 3.5 35B Mixture-of-Experts Transformer with MTP Heads
Merging Technique: MoE Weight Patching & Tensor Grafting
Constituent Models: Methodology: MTP prediction layers stripped of layout corruption and re-grafted cleanly for local engine compatibility.
Merging Technique: MoE Weight Patching & Tensor Grafting
Constituent Models: Methodology: MTP prediction layers stripped of layout corruption and re-grafted cleanly for local engine compatibility.
🚀 3. Technical Enhancements
> 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.
📊 4. Benchmark Competitiveness vs. Frontier Scores
🏆 5. Comprehensive Arena Analytics
> Status: Active Community Benchmarking
// Note: Arena Elo and head-to-head winrates updated continuously as evaluation telemetry processes.🔍 6. SWOT Analysis
> Strengths (S)
- 🛡️ Uncensored Fidelity: Surgically patched to ensure maximum generation throughput without alignment overhead.
- ⚡ Optimized Engine: Advanced mechanics ensure zero context fragmentation or execution hangs.
> Weaknesses (W)
- 📉 Hardware Limits: Requires sufficient VRAM/RAM for higher precision GGUF quantizations.
> Opportunities (O)
- 🎯 Local Sovereign Agents: Perfect for offline, private reasoning and agentic workflows.
> Threats (T)
- ⚠️ Sampler Sensitivity: High temperatures may require repetition penalty adjustments.
⚡ 7. Usage & Deployment Info
> Recommended Settings
- Temperature: 0.2 - 0.7
- Top-P: 0.95
- Backend Engines: Compatible with llama.cpp, vLLM, Ollama, LM Studio, KoboldCPP
⚙️ 8. Backend Compatibility
> Validated Engines:
- [+] llama.cpp: Native support across all quantizations.
- [+] Ollama / LM Studio: Full GGUF compatibility.
📜 9. Disclaimers & Credits
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.
Credits: Gratitude to original base model authors (HauhauCS/Qwen3.5-35B-A3B-Uncensored-HauhauCS-Aggressive) and open-source AI community tools.