Frontier LLM // Resources

Research, docs,
and reference material.

Everything published by Frontier LLM — architecture papers, training reports, API references, and developer documentation — in one place.

Research papers

Design paper · v0.1-alpha · June 2026

EchoNest: A Proposed Hierarchical Recurrent–Attention Architecture for Efficient Instruction-Following

Frontier LLM, Inc.

We propose EchoNest, a compact 102.8M-parameter hierarchical recurrent–attention model that explicitly separates local sequential processing from global context integration across three nested tiers. Designed for efficient inference on commodity CPUs, EchoNest uses recurrent units for local order-sensitive composition and reserves attention for the global tier, trained directly on instruction–response pairs without a separate raw-text pre-training stage.

Download PDF Download Markdown

Parameters 102.8M

Architecture Hierarchical LSTM-attention

Context 512 tokens

Status Training in progress

Formats PDF · Markdown

A full technical report with reproducible specifications and controlled experiments will follow completion of training.

Documentation

Architecture

EchoNest Architecture

Deep-dive into EchoUnit, MicroNest, MacroNest, GlobalEchoCore, and the output projection layer — with component cards and diagrams.

Read docs →