Microsoft-backed G42 scales up Nanda Hindi AI model to 87 billion parameters for India push

The release marks a major upgrade from G42’s initial 13-billion-parameter Nanda model launched in September 2024, as the firm deepens its push into one of the world’s fastest-growing technology markets.

G42, the Abu Dhabi-based artificial intelligence company, rolled out an expanded version of its Hindi-English language model on December 16, heating up competition to capture India's booming digital economy.

The company’s Nanda-87B model contains 87 billion parameters, a measure of an AI system’s complexity and capability, making it the largest open-source model focused on the Hindi language.

The release marks a major upgrade from G42’s initial 13-billion-parameter Nanda model launched in September 2024, as the firm deepens its push into one of the world’s fastest-growing technology markets.

Built by Mohamed bin Zayed University of Artificial Intelligence in partnership with G42's Inception unit and AI chipmaker Cerebras, the model was trained using more than 65 billion Hindi language tokens and uses a specialised tokeniser designed to improve processing efficiency for Hindi text.

Also read: India will likely become a major AI superpower within next four to five years: G42’s Manu Jain

"India deserves world-class technology that speaks its language," G42 India chief executive officer Manu Jain said. The model "opens doors for more inclusive innovation in education, entertainment, enterprise, and beyond".

G42's move comes as global technology giants scramble to capture share in India's AI sector.

The Nanda model handles formal Hindi in Devanagari script as well as casual speech patterns and Hinglish (a mix of Hindi and English used across social media and messaging platforms). It performs tasks including translation, summarisation and text generation.

Inception CEO Ashish Koshy said the upgraded model aims to serve "content creators, educators, developers, and enterprises working across India's digital landscape".

The model has been released publicly on Hugging Face, allowing startups and enterprises to build applications without licensing restrictions. It was trained on Condor Galaxy, a supercomputing system built by G42 and Cerebras.

Richard Morton, executive director of the Institute of Foundation Models at MBZUAI, termed the release as "an important milestone in bringing high-quality, open-access language technology to one of the world's largest and most dynamic linguistic communities".

G42 has been expanding its footprint across emerging markets, developing similar language models for Arabic (JAIS) and Kazakh (SHERKALA) as part of a broader strategy to build AI tools for what the company calls the Global South.

The company's infrastructure includes data centres and cloud computing facilities across multiple countries, positioning it as a regional counterweight to American and Chinese dominance in AI.

Also read: Abu Dhabi’s G42 plans AI-focussed large data centers, supercomputer, Hindi LLM: India CEO

Founded in 2018, G42 has partnerships with major firms including OpenAI, Dell, IBM, Microsoft, Nvidia, Oracle, and Cerebras.

G42 has received investments from global technology investors Mubadala and Silver Lake. In April 2024, Microsoft invested $1.5 billion in G42, and Microsoft vice chair and president Brad Smith joined the G42 board.