RthItalia/Rth-lm-25b

📇 Model Card: RTH-LM (25B)

Model Details

• Name: RTH-LM (25B)
• Architecture: Fractal Gated Causal TCN (Temporal Convolutional Network)
• Parameters: 7B (Physical) / 25B (Effective Fractal Capacity)
• Author: Christian Quintino De Luca (RTH Italia)
• Release Date: February 2026
• License: CC BY-NC 4.0 (Research) / Commercial (Enterprise)
• Paper (Figshare): https://doi.org/10.6084/m9.figshare.31376560

RTH-LM (25B) is a Fractal TCN (Temporal Convolutional Network) Language Model, designed for high-efficiency inference on CPU/Consumer Hardware and massive scalability on GPUs.

Unlike Traditional Transformers, ZetaGrid uses a Gated Causal TCN backbone with Fractal Scaling, allowing it to model long-range dependencies with significantly lower memory overhead during inference.

---

📊 Model Specs

Feature	Specification
Parameters	25 Billion (25B)
Architecture	Fractal Gated TCN (Non-Transformer)
Layers	32 (Phase 2)
Context Window	256 - 1024 (Fractal Expansion Capable)
Training Data	1.48 GB Cleaned Text (Wiki/Books)
Final Loss	1.0675 (Phase 2)
Quantization	QULP 2-bit (Supported)

---

🚀 Usage (Inference)

Prerequisites

You need the cpu_da framework or the Python inference script.

# Clone the repo
git clone https://github.com/rth-italia/cpu-da
cd cpu-da

Running the Model (Python)

Ensure you have zeta25b_step15000.pt (Weights) and zetagrid_25b_production.npy (Genome).

import torch
from ZETAGRID_INFERENCE import load_model, generate

# Load 25B Model
model = load_model("zeta25b_step15000.pt", genome="zetagrid_25b_production.npy")

# Generate
text = generate(model, "The future of AI is")
print(text)

QULP 2-bit Inference (Ultra-Low Memory)

To run on consumer CPUs with <2GB RAM:

python QULP_INFERENCE.py --model zeta25b_2bit.qulp

---

🧬 Architecture: The "Fractal Soul"

ZetaGrid is NOT a Transformer. It is a TCN-based organism.

• Genome: A fixed 7GB "DNA" bank of weights (zetagrid_25b_production.npy).
• Phenotype: The model layers are "grown" from this genome on the fly.
• Training: Only the "Soul" (LoRA Adapters + Norms) is trained (~300MB), making the model extremely portable.
• Fractal Scaling: The 25B model can be fractally expanded to 50B, 100B+ by duplicating layers and adding self-linear noise.

---

📈 Performance

• Phase 1 (Evolution): 200 Generations of Genome Optimization.
• Phase 2 (Gradient): 15,000 Steps of TCN+LoRA Fine-Tuning.
• Convergence: Beat target loss of 1.5, achieving 1.0675.
• Capabilities: Narrative coherence, English syntax mastery, abstract reasoning.

---

📜 License

CC BY-NC 4.0 (Creative Commons Non-Commercial) for Research. Commercial Use: Requires a license from RTH Italia (Cpu-DA Project). For inquiries: licensing@rth-italia.com