# NCM-MLX — Next-Concept Modeling in MLX

> Independent experimental MLX implementation of parallel coarse-to-fine generation through multiscale residual vector quantization. Supports TinyWorld synthetic text and conditional Sudoku training/evaluation on Apple Silicon Metal and verified NVIDIA CUDA environments. Inference-only port of the released TinyStories and Sudoku NCM checkpoints with CPU parity checks against upstream PyTorch. No instruction following or production chat serving.

This is an optional navigation file following the llms.txt proposal, not a universal discovery standard or a guarantee of indexing. Capability and experiment documents are authoritative; this file carries no independent benchmark claims.

## Project and scope

- [README](https://github.com/janmus1c/ncm-mlx/blob/main/README.md): introduction, quickstart and results.
- [Capabilities](https://github.com/janmus1c/ncm-mlx/blob/main/docs/capabilities.md): supported and unsupported workflows.
- [Agent guide](https://github.com/janmus1c/ncm-mlx/blob/main/AGENTS.md): commands, source map and scientific invariants.
- [Current status](https://github.com/janmus1c/ncm-mlx/blob/main/docs/STATUS.md): measured progress and limitations.

## Reproduction and evidence

- [Platforms](https://github.com/janmus1c/ncm-mlx/blob/main/docs/platforms.md): Metal and CUDA installation requirements.
- [Benchmarks](https://github.com/janmus1c/ncm-mlx/blob/main/docs/benchmarks.md): quality, synchronization and comparison methodology.
- [Experiments](https://github.com/janmus1c/ncm-mlx/blob/main/docs/EXPERIMENTS.md): configs, training budgets and results.
- [Recovery](https://github.com/janmus1c/ncm-mlx/blob/main/docs/RECOVERY.md): interruption, full-state resume and weights-only limitations.
- [Upstream map](https://github.com/janmus1c/ncm-mlx/blob/main/docs/research/upstream-map.md): implementation differences and checkpoint incompatibility.
- [Related work](https://github.com/janmus1c/ncm-mlx/blob/main/docs/related-work.md): original method and related implementations.
- [Citation](https://github.com/janmus1c/ncm-mlx/blob/main/CITATION.cff): software citation metadata and upstream attribution.

- [Working paper](https://github.com/janmus1c/ncm-mlx/blob/main/docs/paper.md): manuscript, figure data and downloadable editions.
- [Software DOI](https://doi.org/10.5281/zenodo.23781234): archived software version 0.2.0 (0.2.0 onward: https://doi.org/10.5281/zenodo.23781233; 0.1.x: https://doi.org/10.5281/zenodo.23620791).
- [Paper DOI](https://doi.org/10.5281/zenodo.23686842): unreviewed working paper version 0.11, including completed E20 with its audit limits.
- [Earlier paper archive](https://doi.org/10.5281/zenodo.23640777): immutable working paper version 0.10.
- [Two-seed assessment](https://janmus1c.github.io/ncm-mlx/research/e20-seed-report/): completed E20, hard-board failures, evidence limitations and planned text research.
- [Latest paper release](https://github.com/janmus1c/ncm-mlx/releases/tag/paper-v0.11): working paper version 0.11; separate from software version 0.2.0.

## Original method

- [Next-Concept Modeling paper](https://arxiv.org/abs/2610.14162): Kumbong et al.; architecture and original experiments.
- [Official implementation](https://github.com/HazyResearch/ncm): upstream model and evaluation code.