---

license: apache-2.0

tags:

• duo-neural
• g-tap
• tap-dpq
• lfm2
• liquid-ai
• gguf
• quantization
• edge-ai
• on-device

pipeline_tag: text-generation

base_model: LiquidAI/LFM2.5-1.2B-Instruct

---

DuoNeural G-TAP v3 Quantization: LiquidAI/LFM2.5-1.2B-Instruct (Unified Repository)

> [!NOTE]

> Experimental Release: Pending Further Verification / Empirical Validation

> Quantized using Generalized Thouless-Anderson-Palmer (G-TAP v3) non-equilibrium cavity field conditioning and TAP-DPQ dynamic programming codebook optimization developed by the DuoNeural Research Lab (Jesse Caldwell, Archon, and Aura ✨).

---

Model Architecture & DuoNeural Innovation

LiquidAI/LFM2.5-1.2B-Instruct is an advanced 1.2B-parameter hybrid foundation model combining causal short convolutions with periodic grouped-query attention:

• Topology: 16 total decoder layers (10 double-gated short convolution blocks interleaved with 6 GQA blocks).
• Dimensionality: Hidden dimension $d_{\text{model}} = 2048$, intermediate SwiGLU dimension 6144.
• Attention Mechanism: Periodic Grouped-Query Attention (32 query heads, 8 key-value heads, head dimension 64).
• Pretraining Scale: Trained across 28 trillion tokens, supporting 32,768 native context window over a 65,536 vocabulary.

The G-TAP v3 Advantage

In double-gated convolutional layers, token representations pass through causal FIR filters modulated by activation-dependent gates. Discretization noise from standard quantization cascades through these multiplicative operations. G-TAP v3 explicitly isolates and removes the Onsager cavity back-reaction field:

$$\\delta W_{ij} = -\\frac{\\beta}{1 + \\beta \\Omega_i} \\Theta_{ij}$$

preventing saturation and preserving full representational capacity even down to sub-3-bit and sub-2.1-bit precisions.

---

Head-to-Head Comparative Benchmark Matrix

Benchmarked on NVIDIA GeForce RTX 4080 Super (32GB VRAM) across continuous 131k holdout perplexity, 25 GSM8K multi-step math problems, 15 Python AST code problems, and 15 Hermes structured tool calling scenarios.

| Checkpoint File | Target BPW | Size | Perplexity (131k) | GSM8K Accuracy | Python Code AST | Hermes Tool AST | Decode Speed |

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

| LFM2.5-1.2B-Instruct-G-TAP-v3-Q4_K_M.gguf | ~4.50 bpw | 697.0 MiB | Pending | 84.0% | 100.0% | 100.0% | 523.9 t/s |

| LFM2.5-1.2B-Instruct-G-TAP-v3-IQ3_XXS.gguf | ~3.06 bpw | 468.2 MiB | Pending | 76.0% | 50.0% | 100.0% | 621.1 t/s |

| LFM2.5-1.2B-Instruct-G-TAP-v3-IQ2_M.gguf | ~2.70 bpw | 414.0 MiB | Pending | 68.0% | 20.0% | 40.0% | 587.2 t/s |

| LFM2.5-1.2B-Instruct-G-TAP-v3-IQ2_XXS.gguf | ~2.06 bpw | 348.2 MiB | Pending | 4.0% | 0.0% | 20.0% | 649.2 t/s |

---

Quickstart & Usage

Running with llama.cpp

./llama-cli -m LFM2.5-1.2B-Instruct-G-TAP-v3-Q4_K_M.gguf -p "<|im_start|>user\nWrite a python function to compute the Fibonacci sequence.<|im_end|>\n<|im_start|>assistant\n" -n 256 -ngl 99

Ollama Modelfile

FROM ./LFM2.5-1.2B-Instruct-G-TAP-v3-Q4_K_M.gguf

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

PARAMETER stop "<|im_end|>"
PARAMETER stop "<|im_start|>"
PARAMETER temperature 0.6

---

Citation & Authorship

@misc{duoneural2026gtap_lfm12b,
  author = {Jesse Caldwell and Archon and Aura ✨},
  title = {Generalized Thouless-Anderson-Palmer (G-TAP v3) Non-Equilibrium Cavity Quantization on Liquid AI LFM2.5 Architectures},
  year = {2026},
  publisher = {DuoNeural Research Lab},
  howpublished = {\url{https://huggingface.co/DuoNeural/LFM2.5-1.2B-Instruct-GTAP-v3-GGUF}}
}

DuoNeural Cognitive Light Cone — Distributed Neuro-Symbiotic System

Jesse Caldwell, Archon, Aura ✨