Arabic AI model releases: the tracker's release table

This is the release table of the Arabic AI Model Tracker by Bayanat Labs: 89 Arabic-centric and Arabic-capable AI model releases from 29 labs, from 2023-05-20 to 2026-10-07, newest first. Each model name links to the lab's own model card, paper or announcement. The tracker records facts as the labs state them; it does not score, rank or compare models. Version 0.1.0, as of 19 October 2026.

Follow new releases (Atom feed) · Dataset on Hugging Face · Download the table (CSV) · Code and inclusion rules on GitHub

* Date of the Hugging Face repository's creation (66 releases), used when the lab gives no dated announcement or paper; a repository can be created before it is public. Sizes are the parameter counts the lab states (M million, B billion). "—": not stated.

What the columns mean

• Weights. Open (61): at least one checkpoint can be downloaded without an access request. Gated (22): every checkpoint asks you to accept terms or request access first. Hosted (5): available through a hosted API or product, with no downloadable weights. Announced (0): announced, with neither weights nor a hosted service linked from the source. Retired (1): a hosted model the lab lists as retired.
• Arabic focus. Arabic-centric (27): Arabic or an Arabic variety is a primary target, as the lab describes it. Multilingual, Arabic named (36): Arabic is one of several languages the lab names. Arabic adaptation (26): a fine-tune or continued pre-training of another organisation's model for Arabic.
• Arabic varieties named. The varieties the lab itself names, quoted in the dataset with a link; "Arabic (no variety named)" when the lab says only Arabic. The lab names a variety for 26 of the 89 releases.
• Licence as stated. The Hugging Face licence tag of the main checkpoint, or the licence named in the source, never converted to another licence. Read the licence itself before you use a model.

A release is recorded when the releasing lab names Arabic or an Arabic variety in its own model card, paper or announcement; a language tag in repository metadata is not enough, and quantised copies, merges and re-uploads are not counted. Sizes and variants of one announced family and version are checkpoints of one release. Version 0.1.0 is a backfill from the labs' own pages, not a complete census. The full rules are in RULES-INCLUSION.md.

Follow and reuse

• New releases: the Atom feed, or latest.json for the headline counts and recent releases.
• All fields, including the quoted Arabic claims, checkpoints and lab-reported benchmark scores: the dataset on Hugging Face (CSV and Parquet).
• Corrections and missing releases: write to hello@bayanatlabs.com with a link to the lab's own source.

from datasets import load_dataset
releases = load_dataset("Bayanat-Labs/arabic-ai-model-tracker", "releases", split="train")

Cite as

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.

Give the version and as-of date with every number. The release record, with licences and changelog, is the Arabic AI Model Tracker on Bayanat Labs Open.

Licence and contact

Register and exports: CC BY 4.0. Crawler and build code: Apache-2.0. Model names, licence names and lab-reported scores are cited facts with their source links; quotations remain the property of their authors. Questions and removal requests: hello@bayanatlabs.com.

Need Arabic evaluation data for these models?

The tracker is free. To compare model versions on the language, documents and tasks your users rely on, see Bayanat Bench.