---
title: Cortex-A 0.5 — Live Latent-AR LM
emoji: 🧠
colorFrom: indigo
colorTo: pink
sdk: gradio
sdk_version: 5.50.0
app_file: app.py
pinned: true
license: apache-2.0
short_description: Latent-AR LM trained from scratch, live on CPU
---

# 🧠 Cortex-A 0.5 — live latent-AR playground

This Space serves the **latest weights** of **Cortex-A 0.5**, a ~0.73B two-level language
model pre-trained from scratch in JAX/Flax on a single TPU v5e-8.

- A **deep latent-AR planner** (32 layers over 5-token chunks) predicts the next *chunk
  embedding* through a KV-cached decode.
- A **shallow AR writer** decodes each token conditioned on the **full planner plan**
  (`cond_fuse(concat(token, plan))`), the planner advancing once per 5-token chunk.
- A tied/factorized head reads the tokens out — standard next-token decoding.

The **latent backbone is always on**: its plan conditions every generated token, so what you
see is the real two-level pipeline, not the writer alone. It's a base model — type the start
of a sentence and it **continues** it.