G42, the Abu Dhabi-based technology group, has significantly upgraded its open-source Hindi-English large language model, Llama-3-Nanda (NANDA), now featuring an impressive 87 billion parameters. Collaboratively developed with MBZUAI and Cerebras, NANDA is designed for real-world applications, fluently handling formal Hindi, casual speech, and Hinglish. This model excels in translation, summarization, instruction-following, and transliteration, all while prioritizing safety and cultural alignment.
With a tailored Hindi-centric tokenizer, NANDA enhances training and inference efficiency, making it one of the largest Hindi-focused models available. G42’s CEO for India, Manu Jain, emphasized the importance of providing world-class technology that resonates with local languages, enhancing educational and entrepreneurial opportunities.
This upgrade aims to promote access to AI for over 600 million Hindi speakers and supports India’s rapidly growing digital economy. NANDA 87B is accessible on the MBZUAI Hugging Face page, empowering creators and developers across India’s digital landscape.
Source link
