Fanar

Free, with no paid tier in sight

The FAQ asks “Is Fanar free?” and answers that it is completely free to use. We found no paid plans and no pricing page.

Checked in October 2026 at fanar.qa and the QCRI pages on Hugging Face. The API reference did not load for us (the page says it is served from a public CDN that some networks block), so we could not see rate limits.

Fanar 2.0 turned it into a multimodal platform

According to the release notes, Fanar 2.0 grew the main language model from 9 billion to 27 billion parameters, added step-by-step reasoning shown in Arabic, and lets one prompt trigger several tasks, for example translating an article, summarising it and drafting a social post from the summary. Around the main model sit specialised ones:

• Fanar-Sadiq answers Islamic questions and, according to the site, grounds its answers in authentic sources and cites them.
• Fanar-Oryx generates images and understands images and video, with cultural context in mind.
• Fanar-Shaheen translates between Arabic and English, with dialect support.
• Fanar-Aura handles dialectal speech. The release notes also mention a transcription model for long recordings that separates speakers, and eight new text-to-speech voices in Arabic and English.
• Fanar-2-Diwan writes classical Arabic poetry; its card says it was trained on 118,000 poems and follows the traditional metres and rhyme.

Arabic is the starting point, but check your dialect

The Fanar-2-27B-Instruct card names the dialects it supports: Gulf, Levantine and Egyptian alongside Modern Standard Arabic. That is a more honest list than the usual “all dialects”, and it means Maghrebi speakers should test before relying on it. The same card says the model was built by continuing the training of Google’s Gemma 3 on about 166 billion Arabic and English tokens. That is not a weakness, but anyone planning commercial use of the weights should read Gemma’s terms as well as the Apache licence.

A five-minute test: ask it an everyday question in your own dialect, ask for two lines of verse in a named classical metre and check the scansion, then ask a religious question you already know the answer to and see whether it gives a source you can look up. Put the same three prompts to ChatGPT and the difference, or the lack of one, will be obvious.

Who it suits, and where it falls short

Fanar suits students, teachers, content creators and anyone who wants a capable Arabic assistant for free right now, as well as organisations that prefer a tool built by a research institute in the region. Its limits are on the developer side. The API is not a sign-up-and-ship service: you apply and wait, and there is no published service-level agreement. If you need to run a model inside a closed network, the 27B model is heavy for ordinary machines, and the smaller Jais 3 sizes are a better fit.

For how Fanar compares with Jais and ALLaM, and when each makes sense, see our guide Arabic AI Models: Does a Model That Speaks Arabic Really Understand It? To try it: Fanar’s official site.