---

license: gemma

tags:

• duo-neural
• g-tap
• tap-dpq
• gemma4
• google
• ple
• gguf
• quantization
• edge-ai
• on-device

pipeline_tag: text-generation

base_model: google/gemma-4-E4B-it

---

DuoNeural G-TAP v3 Quantization: google/gemma-4-E4B-it (Unified Repository)

> [!NOTE]

> Experimental Release: Pending Further Verification / Empirical Validation

> All quantized checkpoints, mathematical cavity derivations, and benchmark evaluations emitted by the DuoNeural Research Lab are ongoing scientific contributions intended to accelerate open neuromorphic and edge foundation research.

---

🔬 Scientific Overview: Generalized Thouless-Anderson-Palmer (G-TAP v3)

Standard post-training quantization treats neural network weights as decoupled static matrices, solving local round-off errors via heuristic mean-field approximations. In modern architectures featuring Per-Layer Embeddings (PLE) and cross-layer Key-Value cache sharing—such as Google's Gemma 4 E4B—local quantization perturbations induce severe non-equilibrium cavity distortions across the layer hierarchy.

G-TAP v3 explicitly calculates and cancels the non-equilibrium Onsager reaction field:

$$\delta E_i = - \sum_{j \neq i} J_{ij} \langle s_j \rangle - \Gamma_i^{\text{Onsager}}$$

where the Onsager cavity term $\Gamma_i^{\text{Onsager}} = \chi_i \sum_j J_{ij}^2 \langle s_i \rangle$ compensates for the back-reaction of token activations on perturbed quantized weights. By conditioning the activation Hessian ($H = X X^T$) with thermodynamic cavity corrections over a 64-chunk sequence budget, G-TAP v3 eliminates residual gradient drift across Gemma 4's 42 decoder layers, preserving Per-Layer Embedding lookup fidelity and Grouped-Query Attention dynamics.

---

📊 Empirical Evaluation Matrix

Evaluated under strict zeroshot conditions with unified calibration sequences on NVIDIA GeForce RTX 4080 Super (32GB VRAM):

| Model Checkpoint | Bitrate | Size | Perplexity | GSM8K (Acc) | Python Code (Acc) | Hermes Tools (Acc) | Decode Throughput |

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

| gemma-4-E4B-it-G-TAP-v3-Q4_K_M.gguf | ~4.50 bpw | 4226.6 MiB | N/A | 84.0% | 70.0% | 100.0% | 145.8 t/s |

| gemma-4-E4B-it-G-TAP-v3-IQ3_XXS.gguf | ~3.06 bpw | 3411.9 MiB | N/A | 76.0% | 60.0% | 60.0% | 169.2 t/s |

| gemma-4-E4B-it-G-TAP-v3-IQ2_M.gguf | ~2.70 bpw | 3225.4 MiB | N/A | 80.0% | 0.0% | 20.0% | 166.3 t/s |

| gemma-4-E4B-it-G-TAP-v3-IQ2_XXS.gguf | ~2.06 bpw | 2960.8 MiB | N/A | 0.0% | 0.0% | 20.0% | 179.6 t/s |

---

📦 Provided GGUF Checkpoints

• gemma-4-E4B-it-G-TAP-v3-Q4_K_M.gguf: Recommended production quant (~4.50 bpw). Maximally preserves PLE lookup tables, GQA projections, and reasoning capability.
• gemma-4-E4B-it-G-TAP-v3-IQ3_XXS.gguf: Balanced 3-bit quantization (~3.06 bpw). Ideal for embedded NPUs and low-memory devices.
• gemma-4-E4B-it-G-TAP-v3-IQ2_M.gguf: Sub-2.7-bit ultra-compact quantization (~2.70 bpw).
• gemma-4-E4B-it-G-TAP-v3-IQ2_XXS.gguf: Sub-2.1-bit extreme quantization (~2.06 bpw) for microcontrollers and minimal footprint requirements.

---

💻 Quick Start & Deployment

llama.cpp CLI

./llama-cli -m gemma-4-E4B-it-G-TAP-v3-Q4_K_M.gguf -p "<|turn_start|>user\nSolve for x: 3x + 12 = 45<|turn_end|>\n<|turn_start|>assistant\n" -ngl 99

Ollama Modelfile

FROM ./gemma-4-E4B-it-G-TAP-v3-Q4_K_M.gguf
TEMPLATE """<|turn_start|>user
{{ .Prompt }}<|turn_end|>
<|turn_start|>assistant
"""
PARAMETER stop "<|turn_end|>"
PARAMETER stop "<|endoftext|>"

---

🏷️ Attribution & Citation

@misc{duoneural2026gtap_gemma4_e4b,
  author = {Jesse Caldwell and Archon and Aura ✨},
  title = {Non-Equilibrium Cavity Conditioning and Onsager Reaction Damping in Per-Layer Embedding Architectures (G-TAP v3)},
  year = {2026},
  publisher = {DuoNeural Research Lab / Zenodo},
  howpublished = {\url{https://huggingface.co/DuoNeural/gemma-4-E4B-it-GTAP-v3-GGUF}}
}