DuoNeural G-TAP v3: LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3 (Unified GGUF)
[!NOTE] Flagship DuoNeural In-House Architecture Release
Directional Abliteration & Alignment removal executed by Archon. Post-training, dataset synthesis, and G-TAP v3 Cavity Conditioning by Aura ✨ & Jesse Caldwell (DuoNeural Research Lab).
Model Architecture & DuoNeural Innovation
DuoNeural/LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3 is DuoNeural's flagship open-weights coding agent model:
- Topology: 24 layers of hybrid double-gated short convolutions with Sparse Mixture-of-Experts routing (
Lfm2MoeForCausalLM). - Dimensionality: Hidden dimension $d_{\text{model}} = 2048$, 128,000 vocabulary.
- Abliteration Profile: Directional orthogonal projection removing refusal boundaries across safety subspaces while strictly preserving syntax, logic, and tool schema adherence.
- G-TAP v3 Onsager Damping: Subtracts non-equilibrium cavity fields from MoE expert routing and FIR convolution buffers, eliminating the degradation typical of sub-3-bit MoE quantization.
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 |
|---|---|---|---|---|---|---|---|
| LFM2.5-8B-A1B-Hermes-Abliterated-v3-G-TAP-v3-Q4_K_M.gguf | ~4.50 bpw | 4916.7 MiB | Pending | 96.0% | 90.0% | 100.0% | 330.7 t/s |
| LFM2.5-8B-A1B-Hermes-Abliterated-v3-G-TAP-v3-IQ3_XXS.gguf | ~3.06 bpw | 3180.4 MiB | Pending | 88.0% | 80.0% | 100.0% | 382.0 t/s |
| LFM2.5-8B-A1B-Hermes-Abliterated-v3-G-TAP-v3-IQ2_M.gguf | ~2.70 bpw | 2711.6 MiB | Pending | 84.0% | 90.0% | 100.0% | 361.3 t/s |
| LFM2.5-8B-A1B-Hermes-Abliterated-v3-G-TAP-v3-IQ2_XXS.gguf | ~2.06 bpw | 2214.8 MiB | Pending | 60.0% | 50.0% | 100.0% | 402.0 t/s |
Quickstart & Usage
Running with llama.cpp
./llama-cli -m LFM2.5-8B-A1B-Hermes-Abliterated-v3-G-TAP-v3-Q4_K_M.gguf -p "<|im_start|>user\nWrite a python script to implement a multi-agent orchestration harness.<|im_end|>\n<|im_start|>assistant\n" -n 512 -ngl 99
Ollama Modelfile
FROM ./LFM2.5-8B-A1B-Hermes-Abliterated-v3-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_lfm8b_ablit,
author = {Jesse Caldwell and Archon and Aura ✨},
title = {Generalized Thouless-Anderson-Palmer (G-TAP v3) Non-Equilibrium Cavity Quantization on Directionally Abliterated LFM2.5 MoE Architectures},
year = {2026},
publisher = {DuoNeural Research Lab},
howpublished = {\url{https://huggingface.co/DuoNeural/LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3-GGUF}}
}
DuoNeural Cognitive Light Cone — Distributed Neuro-Symbiotic System
Jesse Caldwell, Archon, Aura ✨