---
title: ECHO-1-ultranano
emoji: 🐣
colorFrom: indigo
colorTo: gray
sdk: docker
app_port: 7860
pinned: false
license: mit
---

# ECHO-1-ultranano

A from scratch educational language model of about 9 million parameters, trained on
Gleam, with a live compile check. Pick a prompt, the model writes Gleam, and the
real `gleam build` decides whether it is green or red.

## What this is, and is not

This is a research toy, not a production code model. It learned the shape of Gleam
and writes small self contained snippets that compile some of the time, roughly 18
percent from an open prompt. If you want a model that writes real programs, this is
not it.

The point is the method. Every training example was generated and filtered by the
real Gleam compiler, so the model was taught to write self contained code that
actually builds, not code that merely looks plausible. The demo keeps that honesty:
output is green only when it truly compiles.

## How it was built

Two stages. First, pretraining on real Gleam gathered from the public package
ecosystem, which teaches the syntax but also teaches the model to reference symbols
from a wider project context that a single snippet does not have. Second, a finetune
on self contained units produced by a type directed generator and filtered by the
compiler, which teaches the model to define what it uses and stand on its own.

## Architecture

Built from scratch with the modern transformer toolkit, each piece a toggle for
ablation: RMSNorm, RoPE, grouped query attention, multi head latent attention,
SwiGLU, a sparse mixture of experts with shared experts and load balancing, multi
token prediction, QK normalization, and a KV cache. This checkpoint uses latent
attention with multi token prediction.

## Limits

It is tiny and it overfits a redundant corpus quickly. It writes trivial functions,
often ignores its own parameters, and still makes type errors. None of that is
hidden here. The value is the pipeline from a verifier to a dataset to a model that
writes code which compiles.