IndiaAI Mission releases new sovereign LLMs to power native language AI

India's Ministry of Electronics and Information Technology has released a new suite of open-source AI foundation models, marking the country's most significant step yet towards sovereign computational capabilities. Developed under the government's ₹10,372 crore IndiaAI Mission, the new native models are built from scratch to perform complex reasoning, agentic tasks, and multimodal processing across all 22 officially recognised Indian languages.

Rather than fine-tuning existing western architectures, the research team trained these baseline models on custom datasets dominated by Indic language text, audio, and visual data. The breakthrough addresses a long-standing efficiency bottleneck in artificial intelligence: conventional global models require up to four times as many tokens to represent Indian scripts compared to English, driving up compute costs and latency. The new native tokeniser reduces that token expansion ratio significantly, allowing faster inference on modest hardware.

How native models compare on Indic benchmarks

Initial benchmarks released alongside the weights show the 70-billion-parameter variant outperforming prominent global open-source baselines on Indic language understanding, translation, and structured reasoning. On the Indic MMLU benchmark, which tests multi-task language comprehension across regional contexts, the native 70B model scored 79.2%, compared to 58.6% for Llama 3.1 70B.

Advanced architecture and agent capabilities

To achieve these gains, the developers implemented mixture-of-experts routing alongside grouped-query attention, allowing the models to maintain low memory footprints during active reasoning. The architecture includes dedicated visual encoder layers, enabling the models to extract text from digitized land records, handwritten documents, and state registry files in various scripts.

Research teams behind the release also incorporated post-training reinforcement learning focused on autonomous task execution. In practical tests, the system demonstrated AI agent functionality by navigating multi-step administrative workflows, querying public databases, and generating bilingual summaries without human intervention.

What makes this release central to India's technological trajectory is its integration with national digital public infrastructure. The open-source weights are being made available to domestic startups, public sector institutions, and researchers through the government's centralised GPU computing cluster, which currently provides access to thousands of high-performance accelerators.

Officials have not disclosed the total financial expenditure for training this specific iteration, nor have they published energy consumption figures for the training runs. Nevertheless, making high-performing multimodal models available without licensing fees provides Indian enterprises with a secure foundation to build local AI applications without exporting sensitive national data.