RTH-LM: A Fractal Temporal Convolutional Language Model

We introduce RTH-LM, a Fractal Gated Causal Temporal Convolutional Network (TCN) for language modeling, designed as an alternative to attention-centric architectures. RTH-LM targets linear-time inference in sequence length and improved data/compute efficiency under constrained training regimes. The model family is organized around a modular separation between a compact shared frozen core (the Genome) and trainable low-rank adapters (the Soul), enabling rapid domain specialization with minimal update artifacts. This paper presents technical specifications, scaling strategies for 120B/1T variants, and initial training signals (Step 15k, Loss≈1.0, PPL≈2.8).

OFFICIAL RESOURCES: Model Hub: https://huggingface.co/RthItalia/Rth-lm-25b Source Code: https://github.com/rthgit/ZetaGrid Organization: RTH Italia (Research & Technology Hub)