Ark-Cortex v3 (ACX-03 · 8B Base-3 Native Ternary 1.58-bit)

---

🏛️ Model Summary

Ark-Cortex v3 (ACX-03) is a sovereign 8.19 Billion parameter large language model operating in native 1.58-bit Base-3 ternary representation (${-1, 0, +1}$), developed and released by the ArkheionNet AI Research Consortium.

ACX-03 is converted from full-precision Qwen3 via a two-stage Quantization-Aware Training (QAT) regime with Orthogonal Trit Gradient Evolution (TGE) and continuous distillation against DeepSeek-R1 reasoning trajectories. Weights across all 36 decoder layers (252 linear projections) are represented as ternary trits scaled by one FP16 factor per group of 64 contiguous weights ($g=64$), matching the official Q2_0 block specification.

🌟 Key Highlights

• Native Base-3 Representation ($g=64$): Weights are exactly represented as $w = s_g \cdot t$, with $t \in {-1, 0, +1}$ and $s_g \in \text{FP16}$ per block of 64 elements, preventing the catastrophic error typical of naive post-training quantization.
• Extreme Memory Footprint Reduction: Shrinks model weight size from 16.38 GB (BF16) down to 4.14 GB (74.7% memory reduction), enabling full 8B model inference on consumer GPUs with 8GB VRAM (e.g., AMD Radeon RX 6600M, NVIDIA RTX 3060/4060) and laptops with 6 GB RAM.
• Deep Reasoning & Chain-of-Thought: Natively generates <think> ... </think> reasoning tokens before synthesizing final answers, preserving complex multi-step deduction capabilities.
• Multi-Token Prediction (MTP): Features 4 chained auxiliary prediction heads ($t+2$ through $t+5$) for enhanced causal foresight and speculative decoding.
• Silicon-Level Optimization: Delivers up to 68 tok/s on consumer AMD RDNA2/RDNA3 hardware via fused HIP Wave32 kernels and sub-warp cooperative execution.

---

📊 Comprehensive Benchmark Evaluation

All evaluations follow standard academic harnesses (lm-evaluation-harness, EvalScope, and llama-perplexity). For reasoning and multi-step tasks, chain-of-thought (CoT) prompting is evaluated using standard greedy decoding (temperature=0.0).

1. Core Reasoning, Math & Knowledge Benchmarks

Summary: ACX-03 retains 91.8% of the Dense Qwen3-8B teacher performance across standard benchmarks while operating with 74.7% less memory footprint.

---

2. Hardware Footprint & Silicon Efficiency

Measured on consumer hardware: AMD Radeon RX 6600M 8GB (gfx1030) and AMD Ryzen 5 5600GT (6 cores / 12 threads):

---

3. Empirical Verification: Chain-of-Thought & Exact Match

Case 1: GSM8K Multi-Step Arithmetic

Prompt: "Natalia sold clips to 48 of her friends in April, and then she sold half as many clips in May. How many clips did Natalia sell altogether in April and May?"

In April, Natalia sold 48 clips.
In May, she sold half of the clips she sold in April, which is 48/2 = 24 clips.
The total number of clips she sold in April and May is 48 + 24 = 72.
The answer is 72.

(Exact match: $72$. Deductive chain, fraction evaluation, and final summation are 100% accurate).

Case 2: Deep Reasoning Chain-of-Thought (<think>)

Prompt: "According to Newton's Second Law of Motion, what is the relationship between force (F), mass (m), and acceleration (a)?"

<|assistant|>
<think>
Okay, let's analyze the question systematically. Newton's Second Law of Motion describes 
the quantitative relationship between the net external force applied to an object, its mass, 
and the resulting acceleration.
Force is defined as the rate of change of linear momentum over time: F = dp/dt.
For a body with constant mass (m), this simplifies to the product of mass and acceleration: F = m * a.
Looking at the options, the correct relation is F = m * a.
</think>
According to Newton's Second Law of Motion, the relationship is:
F = m * a (Force = Mass × Acceleration).

---

🛠️ Quickstart & Serving

1. Using Ollama (1-Line Deployment)

# 1. Download Modelfile and GGUF
curl -O https://huggingface.co/Ooriginador/Ark-Cortex-v3-8B-Ternary-GGUF/raw/main/Modelfile
curl -L -O https://huggingface.co/Ooriginador/Ark-Cortex-v3-8B-Ternary-GGUF/resolve/main/acx3_8b_live_q2_0.gguf

# 2. Build local image
ollama create ark-cortex-v3 -f Modelfile

# 3. Chat with full CoT support
ollama run ark-cortex-v3

---

2. Using llama.cpp CLI & Server

# Run interactive CLI with GPU offload
./llama-cli \
  -m acx3_8b_live_q2_0.gguf \
  -ngl 99 \
  -p "<|im_start|>user\nResolva passo a passo: Se f(x) = 3x^2 + 5x - 7, qual é a derivada f'(x)?<|im_end|>\n<|im_start|>assistant\n<think>\n" \
  -n 512

# Run OpenAI-compatible Local Server
./llama-server \
  -m acx3_8b_live_q2_0.gguf \
  --port 8080 \
  --host 0.0.0.0 \
  -ngl 99 \
  -c 8192

---

3. Using Python (llama-cpp-python)

from llama_cpp import Llama

llm = Llama(
    model_path="acx3_8b_live_q2_0.gguf",
    n_gpu_layers=-1,        # Offload all 36 layers to GPU
    n_ctx=8192,             # Extended context window
    temperature=0.6,        # Recommended for CoT reasoning
    top_p=0.95
)

messages = [
    {"role": "user", "content": "Explain the concept of LIFO in computer science with a code example."}
]

prompt = f"<|im_start|>user\n{messages[0]['content']}<|im_end|>\n<|im_start|>assistant\n<think>\n"
response = llm(prompt, max_tokens=512, stop=["<|im_end|>"])
print(response["choices"][0]["text"])

---

🔬 Architecture & Training Specifications

┌─────────────────────────────────────────────────────────────────────────────┐
│                    ARK-CORTEX v3 (ACX-03) PIPELINE                          │
├─────────────────────────────────────────────────────────────────────────────┤
│                                                                             │
│  [Teacher Oracles] Qwen3-8B + DeepSeek-R1 CoT Distillations                 │
│         │                                                                   │
│         ▼                                                                   │
│  [Two-Stage QAT + Distillation]                                             │
│    • Stage 1: Block-by-Block Decoder Projection Alignment (Block-AP)        │
│    • Stage 2: End-to-End Scale Tuning with Orthogonal Trit Gradients (TGE)  │
│         │                                                                   │
│         ▼                                                                   │
│  [Multi-Token Prediction] 4-Depth Shared Heads (t+2..t+5, Rank 64)          │
│         │                                                                   │
│         ▼                                                                   │
│  [Diamond Lattice sp³ & SWS Tectonic Consolidation (15-18 GPa)]             │
│         │                                                                   │
│         ▼                                                                   │
│  [Sovereign Container Export]                                               │
│    • GGUF Q2_0 (4.14 GB) — llama.cpp / Ollama / LM Studio                  │
│    • .ark (4.34 GB) — Bare-Metal HIP Wave32 / CPU Zero-Copy                 │
│                                                                             │
└─────────────────────────────────────────────────────────────────────────────┘

---

📜 Citation

If you use Ark-Cortex v3 (ACX-03) in your research or applications, please cite:

@article{arkheion2026acx03,
  title   = {Ark-Cortex v3: Sovereign 1.58-Bit Base-3 Ternary Large Language Models with Grouped Scales and Non-Euclidean Hyperbolic Attention},
  author  = {ArkheionNet AI Research Consortium},
  journal = {Arkheion Systems Technical Reports},
  year    = {2026},
  url     = {https://github.com/jhonslife/ArkheionNet}
}

@article{arkheion2026baremetal_ternary,
  title   = {Bare-Metal 1.58-Bit Base-3 Co-Design, Fused Wave32 Silicon Kernels, and Autonomous Training on Consumer Hardware},
  author  = {Arkheion Systems & Hardware Engineering Group},
  journal = {Proceedings of Machine Learning and Systems (MLSys)},
  year    = {2026}
}

---

⚖️ License

The weights and code of Ark-Cortex v3 (ACX-03) are distributed under the Apache License, Version 2.0. Commercial use, academic research, and fine-tuning are fully permitted.