Abu Dhabi’s G42 has made a significant improve to Llama-3-Nanda (NANDA), its open-source Hindi-English giant language mannequin (LLM), which now options 87 billion parameters and units a brand new benchmark in language AI tailor-made for Hindi audio system.
It makes NANDA the most important and some of the succesful Hindi-centric fashions out there in open weights.
Constructed upon Llama-3.1 70B, NANDA 87B has been educated on a curated Hindi-English dataset with over 65 billion Hindi tokens. A customized Hindi-centric tokeniser boosts effectivity, decreasing each coaching and inference time.
NANDA is developed by Mohamed bin Zayed College of Synthetic Intelligence (MBZUAI) in collaboration with Inception, a G42 firm, and Cerebras, makers of the world’s quickest AI inference.
Engineered for real-world use, NANDA is fluent in formal Hindi (Devanagari script), informal speech, and Hinglish. It delivers sturdy efficiency throughout translation, summarisation, instruction-following, and transliteration duties. Security and cultural alignment are core to its design, enabling NANDA to generate context-aware, accountable responses.
NANDA 87B units new benchmark
With over 600 million Hindi audio system and one of many world’s fastest-growing digital economies, India represents a vital marketplace for regional AI innovation. With over 80 per cent of latest web customers preferring native languages, fashions like NANDA can play a pivotal function in bridging digital divides.
Manu Jain, CEO of G42 India, mentioned: “India deserves world-class expertise that speaks its language. NANDA 87B is a significant step in that path, after our first NANDA mannequin was introduced final 12 months.
“As we proceed to scale our operations throughout the nation, this mannequin opens doorways for extra inclusive innovation in training, leisure, enterprise, and past. This improve displays G42’s deep dedication to constructing AI options that serve India’s vibrant AI ecosystem.”
Richard Morton, Govt Director, Institute of Basis Fashions, MBZUAI, added: “At MBZUAI, our mission is to advance AI in ways in which ship broad, optimistic impression for society. NANDA marks an vital milestone in bringing high-quality, open-access language expertise to one of many world’s largest and most dynamic linguistic communities.
“By means of our collaboration with G42 and Cerebras, we’re underscoring the worth of culturally aligned and inclusive AI analysis — work that helps underserved languages and expands entry to superior AI capabilities for a whole bunch of thousands and thousands of Hindi audio system worldwide.”
The mannequin was educated on Condor Galaxy, one of many world’s strongest AI supercomputers for coaching and inferencing, constructed by G42 and Cerebras. A brand new Hindi LLM is now out there as an open-weight mannequin on MBZUAI Hugging Face web page, enabling creators, builders, and enterprises to discover and construct upon its capabilities.
