Science & Technology19 November 20264 min read
IndiaAI Mission launches three language models for Indian-language and offline use
The IndiaAI Mission, under the Ministry of Electronics and Information Technology (MeitY), launched three indigenous foundation language models: EKA-1, Dhir 2B and Dhir 5B.
The models were developed by Indian startups under the IndiaAI Innovation Centre (Foundation Models) pillar, with subsidised access to national AI compute infrastructure. They differ in scale and capabilities, with the Dhir models designed to operate without cloud connectivity.
How the IndiaAI Mission is supporting the models
The IndiaAI Mission was approved in 2024 to build an AI ecosystem through compute infrastructure, indigenous AI development, datasets, skilling and responsible AI. MeitY implements the mission.
The three models were developed by Indian startups under its Foundation Models pillar. The startups received subsidised access to national AI compute infrastructure.
What EKA-1 and the Dhir models are designed to do
EKA-1
Soket Labs Technology and Research Private Limited, also known as Soket AI Labs, developed EKA-1 under Project EKA. It is an open-source sparse hybrid language model with 26.3 billion parameters, trained on roughly 7.5 trillion tokens.
Its sparse design activates only part of the model for a task, conserving computing power. EKA-1 supports all 22 scheduled Indian languages and English. It is intended for mathematics, coding, reasoning and agentic AI, with a focus on banking, financial services and insurance (BFSI) and cybersecurity.
Dhir 2B and Dhir 5B
Bengaluru-based startup ZenteiQ AiTech Innovations developed Dhir 2B and Dhir 5B. These open-source small language models (SLMs) have 2 billion and 5 billion parameters, respectively, and support English and 10 Indian languages.
The models are designed to run offline, on-premise or on-device, including on mobile phones, without cloud connectivity. Their intended uses include document-grounded automation, knowledge retrieval and citizen or enterprise services in defence, government, industry and education. The developer has designed them for low-connectivity or sensitive environments, including border areas.
Why India needs indigenous language models
The stated reasons for developing domestic models include reducing dependence on foreign AI systems in critical sectors, keeping sensitive information local and extending digital services to people who use Indian languages. Lightweight models that can work offline could also support AI use in rural, remote and high-security areas without high-speed internet.
The main concerns and proposed measures include:
- Reducing reliance on foreign AI systems in sectors such as defence, governance and financial services.
- Processing sensitive information on-premise or on-device, so it remains on domestic infrastructure.
- Making services in areas such as healthcare, education, agriculture and e-governance more accessible to non-English speakers, alongside platforms such as BHASHINI.
- Enabling AI tools to function in rural, remote and high-security locations without high-speed internet.
A further challenge is the token tax: English-centric tokenisers can use 35 to 140 tokens to process an Indian-language sentence that requires 10 tokens in English. Since AI services charge per token, this can make regional-language processing slower and more expensive.
The stated need is for culturally representative regional datasets and native Indic tokenisers, developed with native speakers, universities and linguistic institutions. The IndiaAI Mission’s shared platform for datasets, AI models and development tools, AIKosh, is identified as a place to integrate these resources to improve access and support indigenous AI development.
Other indigenous AI models in development
The other named models and suites span language, speech, vision and video generation. Their developers and stated areas of focus include:
| Model or suite | Developer or backing | Type or modality | Stated focus |
|---|---|---|---|
| BharatGen (Param 2 – 17B) | Government-supported consortium led by IIT Bombay under the IndiaAI Mission | Multilingual and multimodal mixture-of-experts large language model | Covers 22 scheduled Indian languages and English; designed for public-sector, governance, education and research applications |
| Sarvam AI, including Sarvam 30B and 105B, Bulbul and Sarvam Vision | Sarvam AI, backed under the IndiaAI Mission | Open-source language models, speech and vision AI | Indic-language processing, document understanding, voice AI and enterprise workloads described as DPDP Act, 2023-compliant |
| Gnani.ai | Gnani.ai, backed under the IndiaAI Mission | Speech-to-speech and conversational voice AI | Indian-language voice recognition, speech synthesis and voice-to-voice interaction for digital assistants and enterprise conversations |
| Krutrim | Ola AI | Multilingual, cloud-native language model | Supports English and 22 Indian languages for consumer, coding, mobility and logistics uses |
| Varya AI | Avataar AI, backed under the IndiaAI Mission | Video generation model | Launched in June 2026 for visual and media content creation |
| Project Indus | Tech Mahindra, as part of the IndiaAI Mission team | Indic language model | Enterprise-oriented foundation model for Indian-language understanding and vernacular applications |