https://janmus1c.github.io/ncm-mlx/feed.xml Software, research articles and working-paper releases for an independent MLX implementation of Next-Concept Modeling. 2026-10-30T15:23:49Z NCM-MLX project https://github.com/janmus1c/ncm-mlx https://janmus1c.github.io/ncm-mlx/research/checkpoint-parity/ 2026-10-30T15:23:49Z 2026-10-30T15:23:49Z Jan Musichttps://orcid.org/0009-0001-5376-7913

Run the released HazyResearch TinyStories and Sudoku checkpoints through an independent MLX port. The article explains conversion rules, tokenizer and numerical parity checks, runnable commands and the limits of the recorded performance comparisons.

https://doi.org/10.5281/zenodo.23781234 2026-10-29T21:39:21Z 2026-10-29T21:39:21Z

Runs HazyResearch's released TinyStories and Sudoku checkpoints natively in MLX on the Apple GPU, inference only. Per-level logits match upstream PyTorch to 3.2e-10 (TinyStories) and 1.2e-13 (Sudoku) in exact fp64; greedy generation is identical. Greedy Sudoku solve rates on all 2,000 boards per difficulty: 99.40 / 96.30 / 86.20% (paper: 99.5 / 96.5 / 86.1%). LM1B, OpenWebText, WMT14 and XSum are not converted.

https://github.com/janmus1c/ncm-mlx/releases/tag/v0.1.1 2026-10-29T21:12:38Z 2026-10-29T21:12:38Z

Packaging and metadata release: pip install ncm-mlx on Apple Silicon or ncm-mlx[cuda] on Linux NVIDIA, author ORCID in citation metadata and a refreshed README. Model code and recorded results are unchanged from 0.1.0. The project's own trained Sudoku and TinyWorld weights are published separately on Hugging Face.

https://doi.org/10.5281/zenodo.23686842 2026-10-27T09:33:29.086196Z 2026-10-27T09:33:29.086196Z

Completed E20 records 97.25 / 79.60 / 33.25% easy/medium/hard solving, compared with seed 0's 98.10 / 83.40 / 40.30% at equal generator exposure. A partial small-model success with a substantial hard-board deficit, two-seed uncertainty and explicit historical audit limits. Controlled text generation is the next research direction, not a validated capability. Unreviewed working paper.

https://doi.org/10.5281/zenodo.23640777 2026-10-25T11:54:17.720827Z 2026-10-25T11:54:17.720827Z

Completed E18, E19 pilot and E19-funded results. The independent MLX implementation solves 98.10 / 83.40 / 40.30% of easy/medium/hard boards. Easy accuracy is within 1.40 points of Original NCM (Kumbong et al.) with a 3.95x smaller generator; medium/hard gaps, additional codec and unmatched budgets remain explicit. Includes checkpoint provenance, uncertainty and A40 timings. Unreviewed working paper.

https://doi.org/10.5281/zenodo.23620914 2026-10-24T15:35:19Z 2026-10-24T15:35:19Z

Coarse-to-Fine Text Generation on a Laptop: A Native MLX Study of Next-Concept Modeling. Version 0.9 corrects reproduction instructions and experimental settings and adds an AI-assistance disclosure. This working paper is not peer reviewed. The method is due to Kumbong et al.; the independent MLX experiments cover synthetic text and Sudoku, with explicit model-size and training-budget qualifications.

https://doi.org/10.5281/zenodo.23620792 2026-10-24T14:45:48Z 2026-10-24T14:45:48Z

Initial MIT-licensed research source release: multiscale residual VQ codec, parallel coarse-to-fine generator, from-scratch training, autoregressive baseline and synchronized benchmarks. Verified execution on Apple Silicon Metal and selected Linux NVIDIA CUDA GPUs. Experimental TinyWorld and Sudoku software; no model weights, general-text validation or upstream checkpoint compatibility are included.