Bayanat-Labs/arabic-ai-model-tracker

Arabic AI Model Tracker

A register of Arabic-centric and Arabic-capable AI model releases since January 2023, with the lab, date, modality, stated sizes, the Arabic varieties the lab names (quoted), lab-reported benchmark scores, licence as stated and weights status. Each row links to its source, with the retrieval time and the SHA-256 of the archived page. Version 0.1.0, as of 2026-10-19. Published by Bayanat Labs (bayanatlabs.com). Code and inclusion rules on GitHub. Arabic summary at the end of this card.

The tracker records facts as the labs state them. It does not score, rank or compare models.

In numbers

Releases	89
Labs and publishers	29
Checkpoints and API model ids	214
Lab-reported benchmark rows	33, for 20 releases
Release dates	2023-05-20 to 2026-10-07
Open weights / gated / hosted access / announced / retired	61 / 22 / 5 / 0 / 1
Arabic-centric / multilingual with Arabic named / Arabic adaptation	27 / 36 / 26
Releases where the lab names an Arabic variety	26

Modality	Releases
Text LLM	46
Speech recognition	11
Vision-language	10
Text to speech	6
Translation	6
Embeddings	5
OCR	4
Dialect identification	1

Version 0.1.0 is a backfill from the labs' own pages, not a complete census. The counts describe the releases recorded so far.

Configs

Config	One row per	Rows
releases (default)	release: an announced model family at one version	89
checkpoints	Hugging Face repository or API model id of a release	214
benchmarks_reported	release, benchmark and metric, as reported by the lab	33

Also in this repository: csv/ (the same three tables as CSV), latest.json (headline counts and recent releases), feed.atom (Atom feed of releases by release date), summary.json (every count on this card), figures/ (the charts and their data), candidates/ (the review queue from the crawler run of 2026-10-19; nothing in it is in the register yet), evidence.csv (URL, retrieval time and SHA-256 of each archived source page), RULES-INCLUSION.md, CITATION.cff and LICENSE.

from datasets import load_dataset

releases = load_dataset("Bayanat-Labs/arabic-ai-model-tracker", "releases", split="train")
open_llms = releases.filter(lambda r: r["modality"] == "text-llm" and r["weights_status"] == "open")

What a release row holds

Field	Meaning
release_id	Stable slug of lab, family and version
model_family, version, previous_release_id	Family and version as announced; link to the previous version's row
lab_name, partners, lab_type, lab_country	Releasing organisation, announced partners, organisation type, country code (null when not recorded)
release_date, date_basis, release_quarter	Date of the lab's announcement or paper, or the Hugging Face repository creation date when neither exists
modality	text-llm, vision-language, ocr-vlm, speech-asr, speech-tts, embedding, translation, dialect-id
arabic_focus	arabic-centric, multilingual-arabic-claimed or arabic-adaptation
arabic_claim_quote, arabic_claim_url	The lab's words naming Arabic (300 characters at most) and where they are
dialects_claimed, dialects_claimed_quote, dialects_claimed_url	Variety codes for the varieties the lab names, with the quote
param_sizes_b, params_active_b	Parameter counts in billions as stated; null when not stated
license_stated, license_url	Licence as stated (Hugging Face licence tag or source text); null when none is stated
weights_status, weights_urls, api_url	open, gated, hosted (API or product, no weights), announced or retired
source_url, source_retrieved_at, source_sha256, evidence	Primary source and each archived source with retrieval time and SHA-256
checkpoints_count, benchmarks_reported_count	Rows in the other two configs
inclusion_reason, inclusion_rules_version, review_status, added_on, last_verified, notes	Provenance and neutral notes

Method

1. Inclusion. A release is recorded when the releasing lab names Arabic or an Arabic variety in its own model card, paper or announcement, in one of the eight modalities, from January 2023. Quantised copies, merges, re-uploads and metadata-only language tags do not count. The full rules are in RULES-INCLUSION.md (version 1).
2. Sources. Each primary source was read once and archived. Fields come from the archived text; each quote was checked by script to appear word for word, and each table score to sit in the stated row and column.
3. Checkpoints. Facts per checkpoint come from a Hugging Face API snapshot: revision, creation date, licence tag, gating, Hugging Face's parameter count from the weight files and the rolling 30-day download count on the snapshot date.
4. Benchmarks. Scores are copied as printed and labelled "as reported by" the lab, with the setting and the place on the page. They are not comparable across rows and are not sorted or ranked.
5. Discovery. A crawler (Apache-2.0, on GitHub) lists new candidates from Hugging Face language tags and an organisation watchlist, the arXiv category feeds and the OpenRouter public models list, filters out quantised copies, merges and adapters, and writes a review queue. A reviewer decides each candidate; nothing is added automatically. The run of 2026-10-19 queued 128 candidates.

The crawler identifies itself as BayanatLabs-OpenData/1.0 (+https://bayanatlabs.com/open; hello@bayanatlabs.com), reads robots.txt, spaces its requests and stops a source on HTTP 401, 403 or 429.

Limitations

• Lab-reported scores use different settings and test sets.
• Dates from repository creation can precede public availability.
• Some releases carry different licences per checkpoint; see the checkpoints config.
• Version 0.1.0 is a backfill and is not complete; some labs and releases are not yet recorded.

Suggest a correction

Send corrections, missing releases or removal requests to hello@bayanatlabs.com with a link to the primary source. Corrections are listed in the changelog on GitHub.

Licence

CC BY 4.0. Credit: "Bayanat Labs Arabic AI Model Tracker, CC BY 4.0, as of 2026-10-19". Model names, licence names and lab-reported scores are cited facts, and quotes are short excerpts with their source links.

Citation

Bayanat Labs (2026). Bayanat Labs Arabic AI Model Tracker, version 0.1.0 (as of 2026-10-19). https://huggingface.co/datasets/Bayanat-Labs/arabic-ai-model-tracker. CC BY 4.0.

@misc{bayanatlabs_arabic_ai_model_tracker_2026,
  title  = {Arabic AI Model Tracker},
  author = {{Bayanat Labs}},
  year   = {2026},
  note   = {Version 0.1.0, as of 2026-10-19},
  url    = {https://huggingface.co/datasets/Bayanat-Labs/arabic-ai-model-tracker}
}

بالعربية

The Arabic text has not yet been reviewed by a native Arabic writer; the English governs.

متتبّع نماذج الذكاء الاصطناعي العربية سجلّ لإصدارات النماذج المتمحورة حول العربية أو الداعمة لها منذ يناير 2023: النماذج اللغوية الكبيرة، ونماذج الرؤية واللغة، والتعرّف الضوئي على النصوص، والتعرّف على الكلام، وتحويل النص إلى كلام، والتضمينات، والترجمة، وتحديد اللهجات. يسجّل كل صف الجهة المطوِّرة وتاريخ الإصدار ونوع النموذج والأحجام المعلنة واللهجات التي تسمّيها الجهة (باقتباس ورابط) ونتائج الاختبارات كما تنشرها الجهة والترخيص كما هو مذكور وحالة الأوزان. يرتبط كل صف بمصدره مع وقت الاسترجاع وبصمة SHA-256 للصفحة المؤرشفة.

يضم الإصدار 0.1.0 (حتى 2026-10-19) عددًا من الإصدارات يبلغ 89، صادرة عن جهات يبلغ عددها 29، وعدد الإصدارات ذات الأوزان المفتوحة منها 61. هذا الإصدار جمعٌ أوّلي من صفحات الجهات نفسها وليس حصرًا شاملًا. لا يقيّم المتتبّع النماذج ولا يرتّبها ولا يقارن بينها.

الترخيص: CC BY 4.0. للتصحيحات أو الإضافات أو طلبات الإزالة: hello@bayanatlabs.com مع رابط المصدر الأصلي.