NCM: The LLM Killer?

Almost every chatbot writes one token at a time, and reasoning models think in tokens you pay for but never see. In October 2026, two preprints posted four days apart tried a different route. Next-Concept Modeling (NCM, Stanford + Together AI) stores text as a stack of coarse-to-fine layers and writes a whole layer per pass. HybridAgents (Stanford, Together AI, UIUC, Meta) lets an agent think silently in hidden-state vectors and switch to text only for tool calls and answers. We go through how both work, the authors' own numbers (Sudoku, summaries, GAIA, cost), where they lose, who can read the silent thoughts, and what you can download today. All figures are the authors' own; both papers are preprints. Sources: - Next-Concept Modeling: Fast, Parallel Coarse-to-Fine Text Generation (arXiv 2610.14162): https://arxiv.org/abs/2610.14162 - NCM code + models: https://github.com/HazyResearch/ncm - Hybrid Reasoning Agents (arXiv 2610.18861): https://arxiv.org/abs/2610.18861 - HybridAgent code + models: https://github.com/togethercomputer/HybridAgent - Google Research on LLM decoding steps: https://research.google/blog/looking-back-at-speculative-decoding/ - OpenAI reasoning guide: https://developers.openai.com/api/docs/guides/reasoning - Coconut (Hao et al., Meta FAIR, 2024): https://arxiv.org/abs/2412.06769 - Jiaru Zou's announcement: https://x.com/Jiaru_Zou - Peter Barnett's reply: https://x.com/peterbarnett_/status/2110511460116005151 0:00 Two broken sentences 0:56 One token at a time 1:23 NCM: write rough, then sharpen 2:27 Speed tests and the Sudoku result 3:31 HybridAgents: thinking without words 4:28 What the agent scored 5:09 Who can read the silent thoughts 5:54 What you can download today