NCM refinement demo

Next-Concept Modeling writes a whole level of latent codes per forward pass, from coarse to fine.

Recorded on M5 Pro with ncm-mlx 0.2.0. Real outputs, not live inference. Every frame below decodes the accumulated latent after a level. Intermediate text can be incoherent; it does not explain model reasoning.

Loading recorded traces…

Independent MLX port; weights and method by HazyResearch / Kumbong et al. (MIT). Not affiliated.