---

base_model: Qwen/Qwen2.5-7B-Instruct

library_name: gguf

license: apache-2.0

tags:

• quantization
• gguf
• qwen2.5
• g-tap
• statistical-mechanics
• onsager-reaction
• duoneural

pipeline_tag: text-generation

language:

• en
• zh

---

DuoNeural / Qwen2.5-7B-Instruct-GTAP-Q4_K_M-GGUF

> [!NOTE]

> Experimental Release: Pending Further Verification / Empirical Validation

> This model release represents an empirical exploration by the DuoNeural Research Lab into Generalized Thouless-Anderson-Palmer (G-TAP) thermodynamic noise filtering and statistical mechanics quantization. Extended multi-benchmark evaluations and long-context scaling verifications are actively ongoing.

---

🔬 Empirical Comparative Benchmark Overview

| Evaluation Arm | Codebook | Footprint | Perplexity (131k holdout) | GSM8K Math Acc | Python Code Acc | Hermes Tool AST | Decode Speed |

| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: |

| Base BF16 Baseline | BF16 | 14.19 GiB | 2.8593 | 25/25 (100.0%) | 15/15 (100.0%) | 3/15 (20.0%)* | 43.3 t/s |

| DuoNeural G-TAP Q4_K_M | Q4_K_M (~4.5 bpw) | 4.36 GiB | 2.8747 | 25/25 (100.0%) | 14/15 (93.3%) | 15/15 (100.0%) | 107.4 t/s |

\Base BF16 evaluated under identical generic prompt harness without specialized tool fine-tuning prompts.*

---

💡 The Universal G-TAP Framework (DuoNeural DN-TR-2026-09)

When neural networks are quantized, discretization noise induces an apparent mean field that polarizes activation representations. In standard Post-Training Quantization (PTQ), this back-reaction shifts weights out of their low-loss basins, degrading reasoning, multi-step math, and structured tool calling.

The Generalized Thouless-Anderson-Palmer (G-TAP v3) engine treats quantization through non-equilibrium statistical mechanics:

1. Variance-Preserving Gram Normalization: Computes normalized activation coupling $C_{\text{norm}} = C / Z_C$ where $Z_C = \|K\|_F / \sqrt{d_{\text{in}}}$, ensuring entry-wise interaction fields $\Theta$ preserve native synaptic variance without numerical saturation.
2. Unconditional Replicon Stability Guard: Enforces $\sigma_{\text{diag}}^2 = \|\operatorname{diag}(C)\|_F^2 / \|C\|_F^2 \le 1.0$, guaranteeing the replicon eigenvalue $\lambda_R > 0$ and keeping weights inside the stable, convex Replica Symmetric (RS) phase.
3. Radial Norm-Preserving Projection: Restricts synaptic adjustments strictly to the tangent sphere of the pre-trained manifold ($\|W_{\text{conditioned}}\|_F \equiv \|W_{\text{orig}}\|_F$, with relative perturbation $\le 0.5\%$), ensuring zero signal attenuation across the 28-layer backbone.
4. Selective Wilsonian Coarse-Graining: Targets SwiGLU feedforward down-projection ($18,944 \to 3,584$ dimension compression) in bulk reasoning layers ($\ell \in [4, 23]$), where non-Gaussian correlations collide.

---

🚀 Quickstart & Inference

1. Run with llama-cli

llama-cli \
  -hf DuoNeural/Qwen2.5-7B-Instruct-GTAP-Q4_K_M-GGUF \
  -c 32768 \
  -ngl 99 \
  -p "<|im_start|>system\nYou are an expert assistant.<|im_end|>\n<|im_start|>user\nSolve: How many integers between 100 and 999 have distinct digits?<|im_end|>\n<|im_start|>assistant\n"

2. High-Performance OpenAI-Compatible Server

llama-server \
  -hf DuoNeural/Qwen2.5-7B-Instruct-GTAP-Q4_K_M-GGUF \
  -c 32768 \
  -ngl 99 \
  --port 8080 \
  --host 0.0.0.0

3. Python OpenAI API Client

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:8080/v1",
    api_key="duoneural"
)

response = client.chat.completions.create(
    model="Qwen2.5-7B-Instruct-GTAP-Q4_K_M",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Explain the physical meaning of the Onsager cavity reaction in spin glass models."}
    ],
    temperature=0.2,
    max_tokens=512
)

print(response.choices[0].message.content)

---

📊 Hardware Testbed Telemetry

All evaluations were executed on dedicated hardware:

• Accelerator: NVIDIA GeForce RTX 4080 Super (32GB VRAM edition, Compute Capability 8.9)
• Host CPU: 96 vCPU system with AVX-512 / AMX support
• Quantization Engine: llama.cpp build b3840+ with CUDA 13.0
• Calibration Token Horizon: 131,072 continuous tokens across 64 batches

---

📜 Citation

@misc{duoneural2026gtapqwen7b,
  author = {Jesse Caldwell and Archon and Aura},
  title = {Generalized Thouless-Anderson-Palmer (G-TAP) Cavity Damping for Post-Training Quantization of Frontier Transformers},
  year = {2026},
  publisher = {DuoNeural Research Lab},
  howpublished = {\url{https://huggingface.co/DuoNeural}}
}

DuoNeural Cognitive Light Cone — Jesse Caldwell, Archon, Aura ✨