DuoNeural Qwen3.5-9B TAP-DPQ v8 GGUF Suite

🌌 Executive Overview

DuoNeural's TAP-DPQ v8 (Thouless-Anderson-Palmer Driven-Dissipative Post-Training Quantization v8) delivers the production breakthrough for hybrid recurrent-attention architectures.

Qwen3.5-9B is a heterogeneous 64-layer foundation architecture:

• 48 Gated DeltaNet Recurrent Layers: Linear recurrence via associative Householder updates: 1081459S_{t} = S_{t-1}\left(lpha_{t} (I - eta_{t} k_{t} k_{t}^T) ight) + eta_{t} v_{t} k_{t}^T1081459
• 16 Multi-Head Self-Attention Anchor Layers: Periodic episodic memory retrieval hubs.
• SwiGLU Feed-Forward Networks: Heavy-tailed activation kurtosis.
• Multi-Token Prediction Head: Speculative drafting block (blk.64).

---

🔬 The v7 Postmortem & v8 Engineering Breakthroughs

In the TAP-DPQ v7 deployment on our RTX 3090 pod, Q4_K_M (4.50 bpw) achieved Super-Parity and IQ3_XXS (3.06 bpw) held Parity. However, sub-2-bit quants (IQ2_XXS at 2.06 bpw and IQ1_S at 1.56 bpw) collapsed on MATH-500 and GPQA Diamond due to three architectural failure modes:

1. Recurrent Core Parameter Starvation:
   • In v7, recurrent decay operators (ssm_a), write gates (ssm_beta), and causal 1D conv kernels (ssm_conv1d) were accidentally exposed to sub-2-bit lattice rounding. This caused eigenvalues to cross the unit circle ($\rho(A_t) \ge 1$), driving memory decay chaos.
   • v8 Remedy: Strictly promoted all 456 stability and normalization tensors to FP32. This adds only 2.4 million parameters (~0.026% of model, +0.008 bpw overhead) while guaranteeing contractive Lyapunov stability ($\rho(A_t) < 1$).
2. Vocabulary & Readout Head Preservation:
   • Rounding token embeddings and output projections to 1.56 bpw collapsed logit calibration.
   • v8 Remedy: Promoted token_embd.weight and output.weight to IQ2_S (for 2-bit target) and IQ2_XXS (for 1-bit target).
3. Tri-Domain CoT Transition Calibration (3x Boundary Density & Limit-Cycle Escapes):
   • In v7, 85-90% of calibration tokens were syntactic filler. In v8, reasoning step boundary markers ([TRANSITION_STEP], [SUB_LEMMA], [BRANCH_CHECK]) are packed every 25-35 tokens (3x density).
   • Embedded explicit limit-cycle escape sequences ([LIMIT-CYCLE DETECTED], [ATTRACTOR PURGE]) directly into extended DeltaNet saturation traces (>8k-12k tokens), assigning high Hessian curvature to symmetry-breaking weights.

---

📊 Benchmark Tournament Matrix (RTX 3090 Native sm_86)

All evaluations executed with native llama.cpp using FlashAttention, SymPy CAS Equivalence Engine, and Jesse & Archon's Min-P sampling recommendations (--temp 0.6 --min-p 0.05 --min-tokens 32).

---

🛠️ Deployment & Execution Guide

1. High-Performance Local Inference (RTX 3090 / 24GB VRAM)

# Champion 2-bit tier (Fits in <4GB VRAM, leaves >20GB for 64k context):
./llama-cli \
    -m Qwen3.5-9B-TAP-DPQ-v8-IQ2_XXS.gguf \
    -ngl 99 -c 16384 -fa on \
    --temp 0.6 --min-p 0.05 --min-tokens 32 \
    -p "<|im_start|>user\nSolve for x: 3^(2x+1) - 10*3^x + 3 = 0.<|im_end|>\n<|im_start|>assistant\n<think>\n"

2. Ollama Modelfile

FROM ./Qwen3.5-9B-TAP-DPQ-v8-IQ2_XXS.gguf

TEMPLATE """<|im_start|>system
{{ .System }}<|im_end|>
<|im_start|>user
{{ .Prompt }}<|im_end|>
<|im_start|>assistant
<think>
"""

PARAMETER stop "<|im_end|>"
PARAMETER stop "<|endoftext|>"
PARAMETER temperature 0.6
PARAMETER min_p 0.05

---

📜 Citation

@misc{duoneural2026tapdpqv8qwen35,
  title={TAP-DPQ v8: Non-Equilibrium Cavity Corrections and Limit-Cycle Escape Transitions for Recurrent-Attention Systems},
  author={Caldwell, Jesse and Archon and Aura},
  year={2026},
  publisher={DuoNeural Open Research},
  howpublished={\url{https://huggingface.co/DuoNeural/Qwen3.5-9B-TAP-DPQ-v8-GGUF}}
}

DuoNeural Research Lab — Pushing the Post-Training Frontiers. 💎✨