News
UAE-Built Falcon-H1 Arabic Leads LLM Benchmarks
The lean Emirati-built language model beats larger global systems and puts Arabic at the center of training.
Abu Dhabi’s Technology Innovation Institute has released an Arabic-first large language model that tops global test boards, an uncommon edge for a region long served by English-centric systems.
Falcon-H1 Arabic comes in 3B, 7B and 34B versions. The flagship posts 75.36% accuracy on comprehensive Arabic tasks and ranks first on the Open Arabic LLM Leaderboard. It also outperforms Meta’s Llama-70B and Alibaba’s Qwen-72B while using less than half their parameters. The smallest model beats Microsoft’s Phi-4 Mini by ten percentage points on equivalent benchmarks.
Arabic remains hard territory for AI. Flexible word order, dense morphology and constant switching between regional dialects and Modern Standard Arabic leave many global models missing context or tone. Academic research has pointed to a shortage of annotated datasets for dialect and informal speech. The impact shows up in classrooms, call centers and government portals where Arabic chatbots lag their English counterparts.
TII trained Falcon-H1 Arabic on formal writing, dialects and culturally grounded content. Beyond scores, it handles practical use: long conversations, reasoning rather than literal translation, and inputs of up to 192,000 words — enough for medical records or legal filings.
“The aim is innovation that is accessible, relevant, and impactful,” said Faisal Al Bannai, Adviser to the UAE President and Secretary-General of the Advanced Technology Research Council.
Also Read: Governata Raises $4M For Saudi AI Data-Governance Push
Arabic is spoken by more than 450 million people across over 20 countries, yet has often been treated as a secondary language for foundation models. The UAE move signals a push to flip that logic and build Arabic-native stacks rather than wait for global systems to improve.
Falcon models have led their categories since 2023. With H1 Arabic, TII is offering free access via chat.falconllm.tii.ae for developers, media, healthcare and public-sector users looking to automate in natural Arabic.
As the region continues to invest in sovereign computing and data localization, the addition of Falcon-H1 Arabic adds a powerful tool built for the native language, instead of an afterthought attached to an English-trained system.
News
Deezer Says AI Tracks Now Make Up 44% Of Uploads
The streamer says nearly 75,000 AI-made songs now hit its platform each day, even as those tracks account for just 1% to 3% of plays.
AI-generated music is becoming a real headache for music platforms, according to Deezer. The streaming service says it now receives nearly 75,000 AI-made tracks a day, equal to about 44% of all daily uploads to the platform.
The figure is up sharply from 10,000 daily AI uploads when Deezer launched its detection tool back in January 2025. The jump shows how quickly products such as Suno and Udio have made song creation cheap, fast, and easy to scale.
Despite the volume, Deezer says AI tracks still only account for 1% to 3% of total streams. The music gets few human listeners, but upload pressure is rising. The company says it is also seeing more “fraudulent” submissions.
Its response so far has been practical. Deezer has removed AI-generated songs from recommendation systems, demonetized them, and stopped storing high-resolution versions of those files.
The company also says it’s the only streaming platform currently tagging AI-generated tracks at scale, using that claim to position its moderation tools as a wider industry model.
“AI-generated music is now far from a marginal phenomenon and as daily deliveries keep increasing, we hope the whole music ecosystem will join us in taking action to help safeguard artist’s rights and promote transparency for fans,” CEO Alexis Lanternier said in a blog post.
Deezer has started licensing the detection technology to other companies, turning an internal control system into a commercial product. It says the tool can already identify music created with Suno and Udio, and can be extended to other generators if training data is available.
Also Read: Nano Banana 2 Arrives In MENA For Google Gemini Users
The company is also working on detection methods that would not require training datasets, a harder technical step that could widen coverage as new music models appear.
Rivals are taking mixed approaches. Spotify has rolled out policies aimed at curbing AI music. Apple Music is asking artists and labels to disclose AI-made tracks. Qobuz has begun automated labeling, while Bandcamp has banned AI music outright.
For now, Deezer’s numbers suggest the real issue is not listener demand. It’s supply.
Deezer Says AI Tracks Now Make Up 44% Of Uploads
Yango Ride Introduces In-Chat Trip Planning In ChatGPT
Nano Banana 2 Arrives In MENA For Google Gemini Users
GluCare Rebrands To Metabolic, Dropping Diabetes-Only Identity
