A ~793M latent-AR language model — a deep planner over 5-token chunks + a shallow AR writer, trained from scratch in JAX/Flax. The latent backbone conditions every token.
Type the start of a sentence and Cortex-A 0.5 continues it.
lower = more focused/confident · 0 = greedy
CPU ≈ 7–8 tok/s · 200 ≈ 25s
0 = off