Google has firmly placed its bet on India. It is stepping in to help the Indian startups develop their own foundational models. At Google I/O India 2025, Manish Gupta, senior director at Google Deepmind, announced that three startups, Sarvam, Soket AI, and Gnani.ai, are building local solutions on top of Gemma family of models.
In addition to startups, Google is collaborating with BharatGen at IIT Bombay to develop Indic language Automatic Speech Recognition (ASR) and Text-to-Speech (TTS) models. These efforts focus on making AI more useful in the Indic context, particularly for underrepresented languages and public sector use cases.
The company said the Gemma 3 series, including the lightweight Gemma 3n variant, is now available for Indian developers with applications ranging from healthcare to speech translation.
“Gemma 3n is designed to enable developers to create multimodal AI experiences that are fast, private and available even offline,” Google said in its blog post. The company added that the model is capable of running on phones and tablets with as little as 2 GB of RAM, making it suitable for a wide range of devices in India.
Google Collaborates with IndiaAI Mission
Google is working directly with Sarvam AI, which has developed a long-form text translation model using Gemma called Sarvam-Translate. It is an open-weight translation model fine-tuned on Gemma3-4B-IT, supporting 22 Indian languages, including Hindi, Tamil, Assamese, and Manipuri.
Unlike generic multilingual models, Sarvam-Translate is optimised for long-form translation across formats such as LaTeX documents, HTML web pages, and OCR-processed files, maintaining structural and syntactic integrity.
“With the help of Gemma, this new model is able to translate long-form and structured documents while preserving context, coherence, and cultural nuance,” the tech giant said in the blog post.
Vivek Raghavan, co-founder of Sarvam AI, on the sidelines of the Google I/O event, told AIM that the company will continue using open-source models available in the market and adapt them for the Indian context. Notably, he added that Sarvam AI is on track to release a model built from scratch under the IndiaAI mission tentatively later this year.
Engineers at Sarvam AI said they plan to surpass the quality of Indic voice models currently offered by its competitors.
Soket AI founder Abhishek Upperwal told AIM that the company has partnered with Google to build a domain-specific agriculture model as part of the IndiaAI mission. The collaboration aims to address on-ground agricultural challenges by leveraging high-quality data and open-source AI models.
“Google will be supporting us in terms of curating the dataset, figuring out domain-specific aspects around it,” said Upperwal. To support this effort, Google is providing compute credits through its startup program and has also committed to financial support. While the exact amount of funding has not been disclosed, Soket AI confirmed that the resources will be used to ensure data quality and model reliability.
Beyond the IndiaAI Mission
Besides the startups selected under the IndiaAI Mission, Google is also working with others, including CoRover, Glance, Entri, InVideo, Nykaa, Dashverse, and Toonsutra.
The tech giant reiterated its support for India’s AI startup ecosystem through programmes like the Google for Startups Accelerator (GFSA), which has supported over 230 Indian startups. The company is expanding its collaboration with the MeitY Startup Hub, launching initiatives such as the ‘Agentic AI Roadshows & Masterclass and the GenAI Startup Buildathon’ in partnership with Nasscom.
Gemma for the Win
Omar Sanseviero, staff AI developer relations engineer at Google DeepMind, announced that the Gemma models have surpassed 180 million downloads. “Thanks to a vibrant community that has created over 80,000 variants based on Gemma. We love how you have embraced Gemma’s multimodality and long context, applying it in powerful ways,” he said.
Sanseviero further added that Google has launched multiple Gemma variants, including Shield Gemma for safe LLM deployment, MedGemma for healthcare applications, Dolphin Gemma for marine research, and Gemma 3n for efficient on-device AI.
Besides, Google also announced that it has made its powerful Gemini 2.5 Flash model available in Indian data centres. Gemini 2.5 Flash processing is now available in India through Single Zone Provisioned Throughput in Google Cloud regions.
“This will especially help developers build for regulated industries like healthcare, banking and finance, and for the public sector, where data residency and low-latency processing are critical,” said Sanseviero.
The post Google Powers India’s AI Mission with Sarvam AI, Soket AI, Gnani.ai appeared first on Analytics India Magazine.
