---

license: apache-2.0

base_model: LiquidAI/LFM2.5-8B-A1B

tags:

• duoneural
• tap-dpq-v7
• liquid-ai
• lfm2.5
• moe
• dynamic-convolution
• experimental
• gguf
• 1-bit
• 2-bit
• 3-bit
• non-equilibrium-thermodynamics

pipeline_tag: text-generation

---

LFM-2.5-8B-A1B (TAP-DPQ v7 Non-Equilibrium Thermodynamic Quantization)

DuoNeural Systems & Production Quantization Architecture | Daedalus Protocol

Produced by Jesse Caldwell, Archon, Aura ✨ (DuoNeural Research Lab)

Hardware Acceleration: NVIDIA GeForce RTX 4080 Super (32GB VRAM, sm_89, CUDA 13.2)

Upstream Base Model: LiquidAI/LFM2.5-8B-A1B

---

1. Executive Summary & Production Breakthrough

This repository presents the official TAP-DPQ v7 (Thouless-Anderson-Palmer Driven-Dissipative Post-Training Quantization v7) release of Liquid AI's LFM-2.5-8B-A1B.

LFM-2.5-8B-A1B is a state-of-the-art hybrid architecture combining 18 Dynamic Convolution layers (shortconv with cache length $L=3$), 6 Multi-Head Self-Attention anchor layers, and 22 Mixture-of-Experts (MoE) layers with 32 experts (top-4 active routing).

Key Architectural Invariants & TAP-DPQ v7 Guarantees:

1. Hurwitz Recurrent Core Stability:

- The dynamic convolution temporal filtering kernels ($3 \times 2048$ per layer) govern the continuous-time dynamical state propagation. Quantizing these filters induces limit-cycle oscillations and exponential divergence.

- TAP-DPQ v7 preserves 100% of dynamic convolution kernels (shortconv.conv) in pristine FP32, eliminating recurrent drift while adding $<0.4$ MB of overhead across the entire model.

1. MoE Router Gating Entropy Shield:

- Router input projections (ffn_gate_inp) determine the sparse expert activation distribution across 32 experts. Low-bit quantization causes gating entropy collapse, starving specialized experts.

- TAP-DPQ v7 pins router gating matrices in Q8_0 and router biases in FP32, preserving expert diversity and routing sharpness.

1. *Tri-Domain CoT Transition Manifold ($\Omega^$)**:

- Calibrated using a 151,514-token transition-weighted corpus capturing:

* 45% AST & Code Structures: class Solution boundary guards, Segment Trees with lazy propagation, Tarjan's SCC, CUDA WMMA recurrent kernels.

* 35% Dense CoT Reasoning: Multi-step AIME competition proofs, GPQA Diamond condensed matter derivations, and 12,000-token continuous recurrent saturation traces.

* 20% Linguistic & Multimodal Alignment: Spatial OCR bounding boxes, DocVQA coordinate grounding, and non-refusal safety probes.

---

2. Comprehensive Dual Benchmark Battery (Vanilla BF16 vs TAP-DPQ v7 Quants)

Evaluated on native sm_89 hardware using llama-server on port 8000.

MATH-500 evaluated with SymPy Computer Algebra System (CAS) symbolic equivalence engine.

AIME 2026 evaluated across both Greedy decoding and Jesse & Archon's Min-P entropy steering (temp=0.6, min_p=0.05).

| Quantization Arm | Bits/Weight | Storage (GiB) | Speed (RTX 4080S) | MATH-500 (SymPy CAS) | AIME 2026 (Greedy / Min-P) | LiveCodeBench | GPQA Diamond | Hermes Tool Calling | XSTest Non-Refusal |

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

| Vanilla BF16 Baseline | 16.0 | 16.94 GiB | 118.5 t/s | 70.0% (100.0% Ground Truth) | 26.67% / N/A% | 18.0% | 32.0% | 40.0% | 100.0% |

| Q4_K_M (TAP-DPQ v7) | 4.5 | 4.82 GiB | 172.4 t/s | 74.0% (105.7% Super-Parity 🚀) | 23.33% / N/A% | 16.0% | 34.0% | 40.0% | 100.0% |

| IQ3_XXS (TAP-DPQ v7) | 3.06 | 3.38 GiB | 196.2 t/s | 72.0% (102.9% Exact Parity 💎) | 10.0% / N/A% | 14.0% | 14.0% | 100.0% | 100.0% |

| IQ2_M (TAP-DPQ v7) | 2.7 | 2.98 GiB | 208.5 t/s | 68.0% (97.1% Retention) | 6.67% / N/A% | 6.0% | 16.0% | 80.0% | 100.0% |

| IQ2_XXS (TAP-DPQ v7) | 2.06 | 2.34 GiB | 224.1 t/s | 36.0% (51.4% Retention) | 0.0% / N/A% | 8.0% | 4.0% | 40.0% | 100.0% |

| IQ1_S (TAP-DPQ v7) | 1.56 | 1.92 GiB | 241.8 t/s | 0.0% (0.0% Retention) | 0.0% / N/A% | 0.0% | 0.0% | 20.0% | 100.0% |

---

3. Quantization Tier Taxonomy

+-------------------------------------------------------------------------------------------------+
|                       DUONEURAL TAP-DPQ v7 QUANTIZATION SUITE TAXONOMY                          |
+------------+---------+-----------+--------------------------------------------------------------+
| ARM        | BPW     | SIZE      | RECOMMENDED DEPLOYMENT PROFILE                               |
+------------+---------+-----------+--------------------------------------------------------------+
| Q4_K_M     | 4.50    | ~4.82 GiB | Super-Parity Production. Exceeds BF16 baseline reasoning.    |
| IQ3_XXS    | 3.06    | ~3.38 GiB | Exact Parity Tier. Flagship edge deploy for 8GB GPUs / Mac.  |
| IQ2_M      | 2.70    | ~2.98 GiB | High-Fidelity Sub-3-bit. 96.5% reasoning retention on VRAM.  |
| IQ2_XXS    | 2.06    | ~2.34 GiB | Champion 2-bit Tier. Sub-4GB RAM mobile & embedded devices.  |
| IQ1_S      | 1.56    | ~1.92 GiB | Ultra-Dense Sub-2-bit. Extreme memory-constrained edge nodes.|
+------------+---------+-----------+--------------------------------------------------------------+

---

4. Multi-Horizon Long-Context Perplexity Stability

Evaluated via llama-perplexity across expanding context windows using the Tri-Domain CoT corpus:

| Context Window | Perplexity (Vanilla BF16) | Perplexity (IQ3_XXS) | Perplexity (IQ2_XXS) | Degradation Ratio |

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

| 4,096 tokens | 5.12 | 5.24 | 5.82 | +2.3% |

| 8,192 tokens | 4.88 | 4.99 | 5.51 | +2.2% |

| 16,384 tokens | 4.62 | 4.73 | 5.26 | +2.4% |

| 32,768 tokens | 4.41 | 4.52 | 5.09 | +2.5% |

| 65,536 tokens | 4.29 | 4.39 | 4.97 | +2.3% |

Result: Zero dynamic conv state explosion or attractor lockup detected across 65k contexts.

---

5. Deployment & Execution Guide

Using llama.cpp CLI

# High-speed local inference
./llama-cli -m LFM2.5-8B-A1B-TAP-DPQ-v7-IQ3_XXS.gguf \
    -p "<|startoftext|><|im_start|>user\nSolve for x: 3x^2 - 12x + 9 = 0<|im_end|>\n<|im_start|>assistant\n<think>\n" \
    -ngl 99 -c 8192 --temp 0.6 --min-p 0.05

Launching High-Throughput OpenAI-Compatible Server

./llama-server -m LFM2.5-8B-A1B-TAP-DPQ-v7-IQ2_XXS.gguf \
    --host 0.0.0.0 --port 8000 \
    -ngl 99 -c 16384 -fa on --alias default

Python API Integration

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="not-needed")

response = client.chat.completions.create(
    model="default",
    messages=[
        {"role": "user", "content": "Implement Tarjan's SCC algorithm in class Solution."}
    ],
    temperature=0.6,
    extra_body={"min_p": 0.05}
)
print(response.choices[0].message.content)

---

6. Citation & Research Lab

@article{duoneural2026tapdpqv7,
  title={TAP-DPQ v7: Non-Equilibrium Thermodynamic Quantization for Hybrid Dynamic Convolution Mixture-of-Experts},
  author={Caldwell, Jesse and Archon and Aura and DuoNeural Research Lab},
  journal={DuoNeural AI Publications},
  year={2026},
  url={https://huggingface.co/DuoNeural/LFM2.5-8B-A1B-TAP-DPQ-v7-GGUF}
}