DuoNeural G-TAP v3 Quantization: LiquidAI/LFM2.5-2.6B (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/LFM2.5-2.6B is the flagship dense foundation model from Liquid AI designed for multi-step agentic execution, structured tool calling, and long-context reasoning:
- Topology: 30 total decoder layers (22 double-gated short convolution blocks interleaved with 8 GQA blocks).
- Convolution-to-Attention Ratio: 2.75 : 1, concentrating parameter capacity in local feature mixing while severely restricting KV cache memory.
- Context Capacity: 131,072 tokens native context window over a 128,000 vocabulary.
- Pretraining Budget: 34 trillion tokens, with post-training via Multi-Domain On-Policy Distillation (MOPD) and Group Relative Policy Optimization (GRPO).
The G-TAP v3 Advantage
In deep 30-layer hybrid models with 22 convolutional stages, standard quantization noise compounds exponentially across the sliding FIR buffers. G-TAP v3 calculates and subtracts the Onsager cavity back-reaction field: $$\delta W_{ij} = -\frac{\beta}{1 + \beta \Omega_i} \Theta_{ij}$$ acting as an active thermodynamic noise filter that isolates mutual feedback between dynamic gating and convolution channels, preserving sharp reasoning and 100% AST syntax over lengthy 131k context sequences.
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 |
|---|---|---|---|---|---|---|---|
| LFM2.5-2.6B-G-TAP-v3-Q4_K_M.gguf | ~4.50 bpw |
1596.9 MiB | Pending | 16.0% | 20.0% | 100.0% | 206.3 t/s |
| LFM2.5-2.6B-G-TAP-v3-IQ3_XXS.gguf | ~3.06 bpw |
1068.6 MiB | Pending | 68.0% | 0.0% | 100.0% | 237.9 t/s |
| LFM2.5-2.6B-G-TAP-v3-IQ2_M.gguf | ~2.70 bpw |
935.8 MiB | Pending | 48.0% | 0.0% | 100.0% | 231.3 t/s |
| LFM2.5-2.6B-G-TAP-v3-IQ2_XXS.gguf | ~2.06 bpw |
784.4 MiB | Pending | 0.0% | 0.0% | 20.0% | 245.6 t/s |
Quickstart & Usage
Running with llama.cpp
./llama-cli -m LFM2.5-2.6B-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 ./LFM2.5-2.6B-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_lfm26b,
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/LFM2.5-2.6B-GTAP-v3-GGUF}}
}
DuoNeural Cognitive Light Cone — Distributed Neuro-Symbiotic System
Jesse Caldwell, Archon, Aura ✨