G42 Unveils NANDA 87B, Empowering 600M Hindi Speakers with Open-Source AI

Abu Dhabi-based technology group G42 has unveiled NANDA 87B, a powerful open-source bilingual large language model designed for Hindi and English. This major release in the generative AI space aims to bridge the digital divide for over 600 million Hindi speakers globally by providing a state-of-the-art tool that understands the nuances of the language.[1][2] The 87-billion-parameter model represents a significant step towards creating more inclusive and culturally aligned artificial intelligence, reflecting a growing trend of developing sophisticated AI tools for languages beyond English. NANDA 87B is being positioned not just as a technological achievement, but as a catalyst for innovation across India's rapidly expanding digital economy, empowering developers, researchers, and enterprises to build new applications for a vast linguistic community.[1][2]

The technical foundation of NANDA 87B is a testament to the rapid advancements in large-scale AI development. Built upon Meta's formidable Llama-3.1 70B architecture, G42 has significantly enhanced its capabilities for the Hindi language.[1][2][3] The model was trained on a massive, curated dataset that includes over 65 billion Hindi tokens, making it one of the largest and most capable Hindi-centric models available with open weights.[2][4][3] This extensive training was conducted on the Condor Galaxy, one of the world's most powerful AI supercomputers, built through a partnership between G42 and the AI hardware company Cerebras.[1][4][3] The development was a collaborative effort, involving G42 subsidiary Inception, Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), and Cerebras.[1][2][4] A key innovation is its custom Hindi-centric tokenizer, which improves efficiency by reducing both the time and computational cost required for training and inference processes.[1][2][4]

The practical applications and performance of NANDA 87B are designed for real-world use. The model demonstrates fluency not only in formal Hindi written in the Devanagari script but also in casual speech and "Hinglish," the widely used hybrid of Hindi and English.[1][2][4] This versatility allows it to deliver strong performance across a range of critical tasks, including translation, summarization, instruction-following, and transliteration.[1][2][4] Core to the model's design are safety and cultural alignment, enabling it to generate responses that are context-aware and responsible.[1][2] By making the model available as an open-weight release on the MBZUAI Hugging Face page, G42 is enabling creators and developers to freely access, explore, and build upon its advanced capabilities, fostering a new wave of innovation.[1][4][3] This move aligns with a broader industry push towards democratizing access to powerful AI tools, which can accelerate progress and broaden the scope of who can contribute to and benefit from AI.

The release of NANDA 87B carries significant implications for India's AI ecosystem and the global push for digital inclusivity. With industry estimates indicating that over 80 percent of new internet users in India prefer local languages, the demand for tailored language models is immense.[2][3] Models like NANDA can play a pivotal role in ensuring that the benefits of AI are accessible to a wider audience, potentially transforming sectors such as education, entertainment, and enterprise services.[1][2] Manu Jain, CEO of G42 India, emphasized this point, stating, "India deserves world-class technology that speaks its language."[1][2][4] The initiative is a clear commitment from G42 to build AI solutions that serve the Global South and address the needs of underrepresented languages in the digital landscape. Ashish Koshy, CEO of Inception, noted that the model is designed to serve a wide range of users, from content creators to enterprises working across India's digital landscape.[1][3]

In conclusion, the launch of NANDA 87B by G42 is a landmark event in the evolution of multilingual artificial intelligence. By building upon a leading architecture and training it extensively on a massive Hindi dataset, G42 has delivered a powerful, culturally attuned, and openly accessible tool. This initiative not only sets a new benchmark for Hindi-English language models but also underscores a strategic commitment to fostering AI innovation within one of the world's fastest-growing digital economies. As developers and researchers begin to leverage its capabilities, NANDA 87B is poised to unlock new possibilities for AI-driven services and applications, ensuring that the future of artificial intelligence is more linguistically diverse and inclusive for millions of people worldwide.