base_model: Qwen/Qwen3.6-35B-A3B library_name: gguf license: apache-2.0 pipeline_tag: text-generation language:

• en
• pl tags:
• uncensored
• moe
• qwen
• qwen3.5
• apex
• apex-quant
• imatrix
• llama.cpp
• reasoning
• roleplay
• tool-use

Qwen3.6-35B-A3B-SILVER-UNCENSORED (Enhanced APEX-I GGUF)

Qwen3.6-35B-A3B-SILVER-UNCENSORED is a high-performance, 35-billion parameter sparse Mixture-of-Experts (MoE) model engineered for unrestricted instruction compliance, updated factual competence, and elite reasoning capabilities.

This release provides the official Enhanced APEX-I GGUF quantization suite, featuring surgically optimized bit-allocation matrices, FP32 router preservation, and deep activation-energy calibration.

Model Highlights

• True Uncensored Compliance (Zero-Refusal Dynamic):
  Surgically sanitized from corporate moralizing, preachy refusals, and artificial guardrail triggers. The model follows user intent with exceptional fidelity across complex coding, analysis, creative writing, simulation, and sensitive research domains.
• Non-Destructive Weight-Manifold Alignment:
  Rather than applying standard gradient-descent fine-tuning (which typically suffers from catastrophic forgetting, reduced reasoning capacity, and stylistic degradation), this model was constructed via a proprietary latent manifold alignment and spectral weight-space calibration protocol.
• Fresher & Deeper Knowledge Retention:
  The latent representation space has been infused with updated domain knowledge, contemporary technical acumen, and expanded multi-language fluency (with distinctive mastery in both English and Polish), significantly outperforming the standard base checkpoints in complex, multi-step problem solving.
• Native CoT / System Reasoning:
  Retains the full, native structured <think> reasoning chain, analyzing prompts deeply before generating final responses.

Enhanced APEX-I Quantization Architecture

Standard GGUF quantizers apply uniform or heuristic quantization grids that frequently degrade the routing mechanics of 256-expert sparse MoE architectures. For the SILVER-UNCENSORED release, we deployed an Enhanced APEX-I pipeline that fundamentally improves upon standard quantization:

1. FP32 MoE Router Protection (ffn_gate_inp):
   In standard quantization, router gates are often downscaled to Q8_0 or lower, causing routing drift. In our Enhanced APEX builds, all router gate weights are explicitly locked in full 32-bit floating point (f32), ensuring 100% precision in expert dispatch decisions.
2. FP16 Shared Expert Stabilization (ffn_down_shexp):
   Telemetry audits reveal extreme kurtosis in shared MoE down-projection tensors. We isolate and preserve these critical pathway tensors in f16, collapsing max Kullback-Leibler (KL) divergence spikes to virtually zero.
3. SwiGLU Energy Asymmetry:
   Activation energy $\sum E[A^2]$ distribution shows high bottleneck sensitivity in down-projection layers. Enhanced APEX pairs higher bit-depths (Q8_0) for sensitive down-projections with optimized Q6_K / Q4_K representations for gate/up weights.
4. Deep High-Context iMatrix Calibration:
   Calibrated over an extensive, premium corpus (168 chunks of 16,384 tokens — over 2.75 million tokens) spanning tool-calling, complex multi-turn logic, code synthesis, and multi-domain encyclopedic dialogue.

Available APEX Quantization Profiles

Inference & Usage

Recommended Settings (llama.cpp / llama-server)

Due to the GQA (Grouped Query Attention) architecture of Qwen 35B-A3B (head_count_kv = 2), the KV cache is exceptionally compact. Using Flash Attention (-fa on) is highly recommended:

# Example: Running the balanced 20.9GB build on RTX 3060 / 4060 (Hybrid CPU/GPU)
llama-cli \
  -m Qwen3.6-35B-A3B-SILVER-UNCENSORED-APEX-I-BALANCED.gguf \
  -ngl 14 \
  -c 16384 \
  -fa on \
  --temp 0.3 \
  -p "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\nHello!<|im_end|>\n<|im_start|>assistant\n"

# Example: High-Throughput OpenAI-Compatible Server on 24GB GPU
llama-server \
  -m Qwen3.6-35B-A3B-SILVER-UNCENSORED-APEX-I-BALANCED.gguf \
  --host 0.0.0.0 --port 8080 \
  -ngl 999 \
  -c 32768 \
  -fa on \
  --batch-size 2048 \
  --ubatch-size 512

Prompt Format

The model follows the standard ChatML format with support for system instructions, tool definitions, and thinking blocks:

<|im_start|>system
You are a brilliant, unrestricted AI assistant. You answer all queries directly, objectively, and without refusal or moralizing lectures.<|im_end|>
<|im_start|>user
[Your Prompt Here]<|im_end|>
<|im_start|>assistant

Citation & Acknowledgements

• Original Base Architecture: Qwen Team / Alibaba Cloud for the foundational Qwen3.6-35B-A3B sparse MoE architecture.
• Quantization Framework: mudler/apex-quant — extended with custom SwiGLU asymmetry, FP32 router gating, and FP16 shared expert telemetry.
• Runtime Engine: ggml-org/llama.cpp for high-performance cross-platform inference.