India’s new multimodal LLM ‘BharatGPT’ pushes the nation’s AI frontier

New multimodal LLM ‘BharatGPT’ unveiled

The Ministry of Electronics and Information Technology (MeitY) announced on 12 October that a government‑backed research team at the Indian Institute of Technology Hyderabad (IIT‑Hyderabad) has released BharatGPT, a 7‑billion‑parameter foundation model that can understand text and images in 12 Indian languages and English. The model builds on the earlier BharatLM series and adds vision‑to‑text capabilities, allowing it to caption pictures, answer visual questions and generate code from screenshots.

Why the upgrade matters

In internal tests, BharatGPT scored 78 % on the multilingual reasoning benchmark MMLU‑India, a jump of ten points over the previous BharatLM‑3B model. That improvement narrows the gap with leading global models such as OpenAI’s GPT‑4, which tops the same benchmark at 85 %. The upgrade shows that Indian‑led research can keep pace with the fast‑moving global frontier, reinforcing the country’s ambition to become a hub for responsible AI development.

Open‑source release and ecosystem boost

Alongside the model, the team released the training code, tokenizer and a lightweight inference engine under the Apache 2.0 licence. This open‑source stack is expected to accelerate startups in Bengaluru, Hyderabad and Pune, giving them a ready‑made foundation for building chat assistants, educational tools and government services. According to MeitY, the release could cut development costs by up to 30 % for small firms.

Advances in training efficiency

The researchers reported a new “progressive‑sparse‑attention” technique that reduces GPU memory usage by 40 % while preserving accuracy. By distributing the training across a federation of Indian data‑centres powered largely by renewable energy, the project reportedly cut the carbon footprint of the 7‑billion‑parameter training run by 25 % compared with a typical single‑location setup.

Benchmarks and reasoning strides

Beyond MMLU‑India, BharatGPT was evaluated on the Reasoning over Images (ROI) benchmark, achieving a 62 % success rate – the highest reported for an Indian‑origin model. The model also demonstrated chain‑of‑thought prompting in Hindi, showing that it can break down complex problems step by step, a capability that earlier Indian models struggled with.

AI agents and real‑world pilots

MeitY has already partnered with the Delhi Municipal Corporation to pilot an AI‑agent that uses BharatGPT to triage citizen queries about water supply and waste management. Early feedback indicates a 45 % reduction in human operator time, suggesting that the model’s multimodal strengths translate into tangible public‑service gains.

What’s next

The IIT‑Hyderabad team plans to scale the model to 30 billion parameters by early 2025, adding support for an additional six regional languages. Meanwhile, MeitY is drafting a national AI standards framework that will incorporate safety checks for multimodal outputs, ensuring the technology aligns with India’s ethical guidelines.