Run Next-Concept Modeling (NCM) on the Apple Silicon GPU with MLX: HazyResearch's released checkpoints with verified parity to PyTorch, plus from-scratch training and benchmarks.
-
Updated
Nov 8, 2026 - Python
Run Next-Concept Modeling (NCM) on the Apple Silicon GPU with MLX: HazyResearch's released checkpoints with verified parity to PyTorch, plus from-scratch training and benchmarks.
Twenty exploratory studies on Next-Concept Modeling and related controls, with code, recorded results and bounded interpretations.
Browser demos of Next-Concept Modeling: small models that draft the whole answer at once, then refine it.
Independent from-scratch implementation of Next-Concept Modeling (NCM, arXiv:2610.14162): a 36M everyday assistant trained on one MacBook
NCM (Next-Concept Modeling) on Apple Silicon/MLX: codec port, surrogate masking, block-seam experiments, plan for modular 1B NCMs
Open research on compute-efficient language models, built on Next-Concept Modeling (NCM). Trained and tested on a single 8 GB GPU.
Taking names and topics off the menu of a Next-Concept Model: forgetting experiments on a 2016 Tesla P40
HyGra: a Windows desktop app for Claude Opus 5.5 via Claude Code on your Pro/Max plan. Fast warm sessions, parallel chats, full history, three themes (HydrarGyrum, Night City, Waifu). Plus a Next-Concept Modeling experiment.
Next-Concept Models on a 6 GB laptop GPU: benchmarks vs autoregressive models and a 3B LLM, a controlled from-scratch comparison on TinyStories, and LoRA fine-tuning
Next-Concept Modeling proves 70B cloud LLMs for basic doc triage are legacy: multiscale discrete concept refinement delivers 100% data sovereignty, sub-.00 marginal unit economics, & verified accuracy on CPU/V100 (91% Sudoku-Hard Best-of-4, 27.93 R-L full 11,334 XSum, zero-pickle safetensors + INT8, FastAPI, clean-room trainers).
Pilot study on level-wise uncertainty and failure modes in Next-Concept Modeling (NCM) using Sudoku-hard puzzles. Investigates entropy probes, early divergence, and forced-prefix interventions.
Can Next-Concept Modeling (NCM) run on a laptop's CPU alone? Timed against GPT-2 and Gemma 4 12B, with a small cache fix that makes its summaries 2.5x faster.
To associate your repository with the next-concept-modeling topic, visit your repo's landing page and select "manage topics."