G42 unveils Nanda 87B upgrade to power Hindi-English LLMs
Built on Llama‑3.1, the open‑weight model targets India’s 60‑crore Hindi speakers with a Hindi‑centric tokenizer and training on Condor Galaxy
Abu Dhabi’s G42 has released Nanda 87B, a major upgrade to its Hindi‑English large language model, positioning the open‑weight system for mainstream use across education, media and enterprise in India.
The model targets tasks from translation and summarisation to instruction following and transliteration, with an emphasis on culturally aligned and safe outputs, according to the company.
What is new in Nanda 87B?
The model scales to 87 billion parameters and has been trained on a curated Hindi‑English corpus with over 65 billion Hindi tokens. A Hindi‑centric tokenizer is designed to improve efficiency during training and inference.
G42 says the system handles formal Hindi, casual speech and Hinglish, and is being released in open weights for developers and enterprises to build on. These details were disclosed by the company and corroborated by Indian business press coverage.
G42 framed the launch as a push to widen AI access for India’s multilingual internet population. The company cites more than 60 crore Hindi speakers and notes that a large share of new internet users prefer local languages, a gap the model aims to bridge with context‑aware responses that respect Indian usage and idioms.
How can Indian developers access Nanda 87B?
According to the company, Nanda 87B is available as open weights via MBZUAI’s presence on a major model‑sharing platform, enabling startups, system integrators and in‑house teams to fine‑tune and deploy for their domains.
Under the hood
Nanda 87B builds on Llama‑3.1 70B and was trained by researchers at MBZUAI with G42’s Inception unit and Cerebras. Training ran on Condor Galaxy, a large G42‑Cerebras supercomputer for AI workloads.
The companies say the model shows strong performance in core Hindi‑English tasks, including code‑mixed inputs common in daily use.
Manu Kumar Jain, CEO of G42 India, has positioned the upgrade as part of a broader plan to localise advanced AI for the country’s developers and consumers.
For global context, G42’s Group CEO Peng Xiao oversees the group’s AI strategy and international expansion through its portfolio companies.
The first public Nanda model, a smaller Hindi‑centric system introduced in September 2024, set the stage for this larger release.
In the broader landscape, Meta’s Llama‑3.1 series, on which Nanda 87B is based, natively supports eight languages including Hindi and offers long‑context capabilities, providing a strong multilingual foundation for Hindi‑English applications.