h9-tec/arabic-ai-atlas reorganized its data again in the 2026-11-04 nightly auto-refresh commit (afd48cc, "chore: nightly refresh 2026-11-04"). This repository is an index that gathers Arabic AI models and datasets by country and shows them as a map and lists, and it is offered as a skill that agents can install; the license is MIT. This commit is not a new feature but the work of re-extracting the figures read from Hugging Face and regenerating the tables and JSON.
The focus of the change is the ⬇ column (download count) in the ASR (speech recognition) table and the resulting rankings. wav2vec2-large-xlsr-53-arabic dropped from 2.4M to 2.2M, and Whisper Quran rose from 307K to 329K, pushing out SeamlessM4T v2 and taking the 4th place. SeamlessM4T v2 fell from 309K to 307K, MMS-1b-all from 276K to 274K, wav2vec2 quran phonetics from 224K to 236K, whisper-large-v3-turbo-ar-quran from 39K to 36K, and wav2vec2 large xlsr 53 arabic quran v final from 12K to 11K. QwenCleo-ASR rose from 45K to 46K, and the items tied at 3K were rearranged in the order NVIDIA FastConformer Arabic, Audar ASR V1 Turbo, egyptian-arabic-wav2vec2-xlsr-53. The README's refresh date badge changed from 2026-11-03 to 2026-11-04, and the item count badge stayed at 3302, so no new items were added.
The scope of the regeneration is broad. Not only the single README page, but also dist/atlas.json (6705 lines added and 6705 deleted), dist/llms.txt, the table documents in docs/tables including asr, benchmark, and dataset, the map images assets/map.svg, geo.svg, tree.svg, and the Hugging Face response cache data/.cache/hf.json were all modified together. The total change volume of the commit is 14958 lines (7479 added, 7479 deleted). The README has 174 ASR items and 1163 dataset items, and exposes 20 of each. In the dataset table, prophet mosque library remained as 10K–100K rows, mit, 386K, and Waqfeya Library is listed as 10K-100K books, mit, 131K.
However, what this commit leaves behind is nothing more than a single day's fluctuation in the download figures. There are no changes to model weights, licenses, or performance scores, and no new evaluation figures were added to the README tables either. The new value of the changed Waqfeya Library row on the dataset side cannot be confirmed from this commit information alone. Because it is a snapshot repeated every day, it is hard to use as a basis for judging the superiority of a particular model, and it is closer to a means of surveying Arabic AI assets or narrowing the range of agent exploration.