Shanghai Team Open-Sources NCP: 8.9B Model Predicts Concepts Instead of Tokens

A Shanghai team proposed and open-sourced NCP-ArchPreview, an 8.9B latent-space language model that predicts discrete multi-token concepts instead of next tokens, matching a 7B model's pretraining loss with only 51% of the training compute.

2026-09-13 ~ 2026-09-14 · 2 related posts

• NCP: latent-space LM matches OLMo-3-7B pretraining loss with 51% of tokens — mark_k · 2026-09-13
• Shanghai team open-sources NCP-ArchPreview, an 8.9B model that predicts concepts instead of tokens — wonderwomancode · 2026-09-14