Google woos India's booming AI developer community with new tools, access to latest models

Representative image

Google aims to tap India's booming artificial intelligence (AI) developer community by introducing a suite of tools, programmes, and partnerships. The move is aimed at helping developers in building AI-driven products and solutions for both domestic and global markets.

"India is at the cornerstone of our global AI mission. With its large mobile-first population, booming startup ecosystem, and diverse linguistic landscape, we're uniquely positioned to drive the AI innovation globally," Seshu Ajjarapu, Senior Director, Google DeepMind told Moneycontrol in an interview.

Ajjarapu said that India is at the forefront of adopting the firm's Gemini family of AI models, which are used by over 1.5 million developers worldwide. He added that India accounts for one of the largest user bases of Google's online developer platform, Google AI Studio, designed for rapid prototyping and experimentation with generative AI models.

Manish Gupta, Director at Google DeepMind, said the tech giant is currently focusing on three areas - multimodal, multilingual, and mobile - which it views as key AI opportunities. "We have been working together with our colleagues globally to infuse all these capabilities into Gemini," he said.

Wider access to AI models

On July 17, at a company's developer event in Bengaluru, Google said it is releasing Gemma 2, the next generation of its open source AI model, to all developers in India.

First introduced in Google I/O developer conference in the United States in May, Gemma 2 features a new architecture for better performance and efficiency and will be available in both 9 billion and 27 billion parameter sizes.

Gemma's tokenizer, which breaks down text into smaller units for AI processing, is particularly powerful for building multilingual solutions that understand and respond to India's diverse languages, the company said.

This was demonstrated by Navarasa, a multilingual variant for Indian languages built on Gemma by Telugu LLM Labs, a joint initiative by LlamaIndex's Ravi Theja Desetty and Knownwell's Ramsri Goutham Golla. It aims to provide better AI experience for Telugu speakers worldwide. Navarasa currently supports understanding 15 Indian languages.

Google is also making a 2 million token context window available on its flagship Gemini 1.5 Pro model to all developers in India, after initially launching it through a waitlist in May.

Context window size determines how much data (words, images, videos, audio or code) a model can process at once. This essentially means that the bigger a model's context window, the more information it can take in and process in a given prompt. For instance, a 1 million token window can process up to 1 hour of video, 11 hours of audio, or extensive codebases and text of up to 1,500 pages.

Benchmarking Indian language LLMs

Google is also introducing IndicGenBench, a multilingual benchmark suite designed specifically for Indian languages. This is aimed at helping developers build high-quality language models that can accurately represent India's linguistic diversity, the company said.

Developed by Google DeepMind’s India unit, the benchmark suite can be used to evaluate language generation capabilities of large language models (LLMs) across diverse user-facing tasks in 29 Indian languages spanning 13 writing scripts and four language families.

Among the supported languages are Hindi, Kannada, Bengali, Gujarati, Tamil, Telugu, Malayalam, and Marathi, as well as underrepresented languages such as Manipuri, Maithili, Konkani, Marwari, and Bodo.

"For many of the languages, this is the first such benchmark, which will spur more innovation," Gupta said.

The tech giant is also open-sourcing a new technology called CALM (Composition of Language Models) that allows developers to combine their specialised language models with Gemma models. The technology, developed by Google DeepMind's India unit, enables developers to create powerful, efficient, and nuanced solutions that cater to specific use cases and linguistic variations, the company said.

For example, if a developer is building a coding assistant in English, they can also provide coding assistance in Kannada by using a specialist model in CALM.

Gupta said the initial motivation to develop this technology was to enhance language inclusivity in AI models.

"Our team developed a small model called Morni, which understood Indian languages very well. We wanted to combine it with the power of Gemini, a much richer model with a deeper understanding of the world, but not as good in understanding Indian languages. So how do you combine the two to get the best of both worlds?" he said.

The DeepMind India team has also played a major role in developing a new framework called Matformer framework, which would enhance on-device AI capabilities, Gupta said.

The framework will be available in the second version of Gemini Nano, which is expected to be released soon. It will allow developers to mix and match different-sized Gemini models within a single framework, optimising for both high performance and low resource consumption. This is expected to translate to smoother, faster, and more accurate AI experiences directly on users' phones.

"Developers can then choose: If I want the highest quality, I use the largest model. If I want to preserve battery life, and a small model will be good enough. I can just choose the smaller model without having to deploy all of these different versions of Gemini onto one device" Gupta said.

New APIs and speech data

Project Vaani, a collaboration between Google, the Indian Institute of Science (IISc), and ARTPARK (Artificial Intelligence & Robotics Technology Park), has also completed its first phase, the company said. The project will provide developers with over 14,000 hours of speech data across 58 languages, collected from 80,000 speakers in 80 districts.

First announced in December 2022, Project Vaani aims to collect and transcribe open-source anonymised speech data from all 773 districts of India, ensuring linguistic, educational, urban-rural, age, and gender diversity in three phases. The first phase focuses on 80 districts across 10 states.

Gupta said they are now in the middle of phase two that will cover all states in India spanning 160 districts.

Apart from these announcements, Google said it will soon launch a Agricultural Landscape Understanding (ALU) Research API in limited preview, with a goal to make agricultural practices more data-driven and efficient.

The API aims to address various challenges faced by farmers, such as accessing subsidies and capital to improve yields and market access. It will leverage AI and remote sensing to map individual farm fields across India, potentially offering landscape insights at the farm field level rather than at the current aggregate level.

In a blog post, Google said the API was built on Google Cloud and the company's extensive research, including collaborations with the Anthro Krishi team and India's digital AgriStack. Early select partners such as Ninjacart, Skymet, Team-Up, IIT Bombay, and the Government of India are currently exploring the usage of ALU information, it said.

Google has also introduced India-specific pricing for developers using Google Maps platform, which the company claims is up to 70 percent lower on most APIs. The firm has also collaborated with the Open Network for Digital Commerce (ONDC), to offer developers building for ONDC up to 90 percent off on select Google Maps Platform APIs, it said.