Fanar 2.0
Qatar's sovereign Arabic model — dialect-aware, free, open-weight
Overview
Fanar 2.0 is Qatar's Arabic-centric generative AI stack from QCRI at Hamad Bin Khalifa University. Built on a Gemma-3-27B backbone, it handles Modern Standard Arabic plus Gulf, Levantine and Egyptian dialects, with a free chatbot, mobile apps, open weights and a developer API.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 13 Jul 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
What you can produce with Fanar 2.0
- Free Arabic web chatbot + iOS/Android apps
- Open weights on Hugging Face (9B and 27B)
- Developer API (access-request gated)
- Dialect support: Gulf, Levantine, Egyptian + MSA
- Multimodal: image generation, speech, Arabic poetry
- Data and inference hosted entirely in Qatar (sovereignty)
ASEAN Perspective
Fanar 2.0 in Southeast Asia
ASEAN-region availability and pricing notes coming soon. Drop the editorial team a note via /contact/ if you can supply local context (Singapore/Malaysia/Indonesia/Thailand/Vietnam).
Where Fanar earns its place is dialectal Arabic and cultural grounding — it tops Belebele and AraDiCE dialect benchmarks and the Arabic Cultural Value Alignment test, beating same-size rivals in Gulf, Levantine and Egyptian Arabic. The whole stack runs inside Qatar, which is the actual selling point for government and regulated buyers who can't ship data to a US provider. Reality check: at 27B parameters it's a mid-size model, not a GPT-5 or Claude competitor for general English reasoning, and larger sovereign models like Jais-2-70B outscore it once you leave its weight class. The free web chat and iOS/Android apps are open to anyone; weights sit on Hugging Face for self-hosting; the API needs an access request rather than instant signup. Best fit: developers and institutions across the Gulf who need trustworthy dialect-aware Arabic with data kept in-country, not a general-purpose frontier model.
What people say
Coverage of Fanar sits mostly in academic papers and regional tech press rather than G2 or Trustpilot, which fits a state-backed research release more than a commercial SaaS product.
The Fanar 2.0 technical report (arXiv 2603.16397) is the meatiest source. It states the 27B instruct model was continually pre-trained from Gemma-3-27B on roughly 120B curated tokens, and that despite using about 8x fewer pre-training tokens than Fanar 1.0 it still gained about 3.5 points on dialect benchmarks. Evaluators found it best-in-class within its size range on Modern Standard Arabic, dialectal Arabic and Arabic-knowledge tasks, while conceding that bigger models — Karnak 40B, Llama-3.3-70B, Jais-2-70B — pull ahead on raw capability once you allow more parameters.
The dialect and culture results are the recurring highlight. Fanar posts the best scores among evaluated models on Belebele and AraDiCE's manual dialect translation of PIQA, plus the Arabic Cultural Value Alignment benchmark. QCRI also leaned on human feedback: over 300 testers from across the Arab world rated outputs, and the team used Direct Preference Optimization to steer answers toward regional cultural and ethical norms. Against Jais, Fanar's own data-filtering recipe added about four points on Arabic HellaSwag versus a Jais-style filter, and the team notes Jais-filtered training plateaued earlier.
On access, the picture is genuinely open. The fanar.qa web chat and the iOS/Android apps are free; weights for Fanar-1-9B and Fanar-2-27B-Instruct are published on Hugging Face under QCRI's account; and third-party hosts like Featherless offer flat-rate serverless inference from around $10/month for teams that don't want to run their own GPUs. The first-party developer API is free but gated behind an access-request form, with a three-month free-access offer floated for some communities.
What's missing is independent, at-scale user sentiment — there's little public Reddit or enterprise-review chatter, so day-to-day reliability, latency and support quality are hard to judge from outside. Work on Fanar 3.0 is already underway with a December 2026 target, so the current 27B flagship reads as a waypoint rather than the destination.
Summary of public user & expert reviews, compiled by RECATOOLS.
About this listing
This entry was compiled from publicly available data including Fanar 2.0's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with Fanar 2.0 unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to Fanar 2.0 directly →
Spotted something out of date? Suggest an update →
Alternatives to Fanar 2.0
More in LLMs & Chat
