IIT Madras and NVIDIA launch four open Indic-language AI models

This launch establishes a shared technical layer for Indian edtech, moving away from fragmented, proprietary APIs toward a unified Digital Public Infrastructure. By providing open-weight models, India ensures that critical educational technology remains under domestic institutional control rather than relying on foreign commercial vendors. For the broader ecosystem, this move pressures private edtech firms to innovate at the application layer rather than charging for basic model access. NVIDIA’s participation highlights a strategic pivot toward embedding its Nemotron stack within national sovereign AI projects globally. Watch for independent benchmark scores to emerge, which will determine if these open models can truly compete with closed-source alternatives like Google's Speech-to-Text.

NVIDIA has been systematically embedding its NeMo and Nemotron frameworks into national AI programs across Asia, positioning the company as infrastructure provider rather than just a chip vendor. In March 2025, NVIDIA announced a partnership with India's Ministry of Electronics and Information Technology to build AI infrastructure supporting 10+ Indian languages, a move that laid the groundwork for projects like the Bharat EduAI Stack. The company has replicated this pattern in other markets: NVIDIA signed a sovereign AI agreement with Indonesia's government in early 2025 to develop Bahasa-language models using the same NeMo toolkit, signaling that the Bodhan AI collaboration is part of a deliberate multi-country strategy rather than a one-off academic project. The regulatory and policy environment in India has accelerated demand for open-weight Indic-language models. India's National Education Policy 2020 mandated multilingual instruction through grade 12, creating a structural need for AI tools that handle low-resource languages at scale. The Indian government's Digital India Bhashini initiative, which aims to provide real-time translation across 22 scheduled languages, has processed over 1 billion API calls since its 2022 launch, demonstrating the scale of demand that Bodhan AI's models are designed to serve. Meanwhile, AI4Bharat, the IIT Madras research group behind several of the underlying Indic models, published benchmark results showing its IndicTrans2 model outperformed Google Translate on 8 of 11 Indic language pairs, providing independent validation of the open-source approach that the Bharat EduAI Stack builds upon. On the technical side, the models released under the Bharat EduAI Stack target specific pipeline stages rather than general-purpose generation, a design choice that reflects lessons from production deployments. NVIDIA's Nemotron-4 340B model, which serves as the base architecture for several of these Indic adaptations, achieved state-of-the-art results on multilingual reasoning benchmarks when released in mid-2024, though the distilled task-specific versions used in the Bharat stack trade breadth for latency and cost efficiency. IIT Madras's AI4Bharat team previously demonstrated that Indic OCR models trained on NVIDIA hardware achieved 98.2% character accuracy on Devanagari script, outperforming commercial alternatives from Google Cloud Vision and AWS Textract on Indic scripts. These results suggest the open models can compete with proprietary alternatives on the specific tasks that matter most for real-time corporate training interactions.