FRONTIER LLM  //  MODEL CARD

EchoNest

v0.1-alpha Production · fine-tune in progress Early access

EchoNest 3 is a compact ~56M-parameter transformer language model, trained CPU-only on ~246M tokens of story text and in production since Nov 2, 2026. A chat / code / tool-use fine-tune is in progress. Designed to run on commodity hardware without a GPU.

AT A GLANCE
Parameters~56M (EchoNest 3)
Context window—
ArchitectureTransformer
(decoder-only)
Progresswaiting…
LicenseEarly access
In production · since Nov 2, 2026
EchoNest 3 (~56M transformer)
EchoNest 3 replaces EchoNest v2 (12M). On the same held-out story text it reaches perplexity 4.57 vs v2’s 9.92. Trained CPU-only on ~246M tokens.
Sample: “The brave knight entered the dark forest and began to search for something special. He was determined to find it! Suddenly, he saw a large tree…”
Next: a chat / code / tool-use fine-tune is in progress (not live yet), ETA around Nov 3. The earlier 1.26B experiment has finished and is retired — it did not beat production. Track progress on Status · rails format on Agentic Rails.
01

Overview

EchoNest is designed for coding assistance, conversation, short chain-of-thought, and agentic tool rails — with an emphasis on efficient training and inference on commodity CPU hardware.

Earlier EchoNest versions used a three-tier nested recurrent structure (shown below). The current production model, EchoNest 3, is a compact decoder-only transformer — still lightweight and practical to run without dedicated accelerators.

THREE-TIER NESTED ARCHITECTURE
MacroNest · global context integration
MicroNest · mid-range composition
EchoUnit · local sequential processing
02

Model details

Parameters~56M (EchoNest 3)
ArchitectureDecoder-only transformer
Training objectiveCausal LM on story text
03

Training data

EchoNest 3 was pretrained CPU-only on ~246M tokens of story text so it learns coherent English first. Chat, code and tool-use data are being added as a separate fine-tune (in progress).

DATASETEXAMPLESCOVERAGE
TinyStories~246M tokLanguage base (EchoNest 3)
Stanford Alpacain progressChat / code / tool-use fine-tune
Python Code Instructionsin progressChat / code / tool-use fine-tune
Current cap~246M toklanguage pretraining, CPU-only

An earlier 1.26B experiment (mixed data from scratch) finished but did not beat production and is retired. EchoNest 3 instead builds a strong language base first, then fine-tunes.

04

Training

Training tokens~246M (language base)
HardwareCPU only (this box)
05

Evaluation

EchoNest 3 vs EchoNest v2 on the same held-out story text (lower perplexity is better).

4.57
EchoNest 3 perplexity
9.92
EchoNest v2 perplexity
~56M
EchoNest 3 parameters
06

Intended use

→Python code generation from natural-language descriptions
→Instruction-following tasks in English
→Lightweight deployment where a small, fast model is preferred over a large general-purpose one
07

Limitations

Context length
128 tokens on this CPU run. The model cannot reason over long documents or multi-turn conversations exceeding this window.
Language
English only.
Code languages
Primarily Python in v0.1. Multilingual code support is planned for a future release.
Scale
At 15.82M parameters this CPU run is sized to be iterable. The three-tier nest is unchanged; a larger nest can return after language is coherent.
Early stage
v0.1-alpha is a research release. Outputs should be reviewed before use in production.
08

Roadmap

v0.1-alpha15.82M params · 128-token context · TinyStories CPU language pretrainTraining
v0.2Evaluation suite, inference optimisation, public weightsPlanned
v1.0Safety review, Agentic Rails integration, API releasePlanned

Build on EchoNest
while it's still early.

Join the early-access program to test the alpha, shape the roadmap, and get first access to the v1.0 API.

Get API access Talk to sales