DeepMind's India team contributed significantly to Gemini 2.5 Flash AI model, says Google's Manish Gupta

The model builds on Gemini 2.0 Flash, which was introduced in December 2024 and made available in February.

Gupta said Gemini 2.0 Flash model gives 5X higher performance per dollar than DeepSeek’s R1.

“DeepSeek is seen as a company that brought a paradigm shift in cost efficiency, that's not quite true. It was very impressive no doubt... People think DeepSeek came up with some amazing breakthrough on efficiency. Flash is actually more efficient.” Gupta said.

Gupta’s comments come after Alphabet CEO Sundar Pichai during the company's Q4 earnings call in February said the search giant's Gemini Flash 2.0 and Flash Thinking 2.0 models were “some of the most efficient models” available, outperforming even DeepSeek's R1 and V3.

Pichai's remarks came after the Chinese AI lab DeepSeek triggered a Wall Street meltdown in late January with its claims of developing a model those rivalling top-tier models from American companies such as OpenAI, Meta and Google at a fraction of the cost.

The claims made investors question the billions of dollars being poured in by tech companies to develop their AI models and products.

Developing LLM benchmarks

Gupta’s team in India has also been central to contributing to language technologies, especially for Gemini 2.0, where the team evaluated a set of 29 Indian languages through a benchmark.

IndicGenBench, released in 2024, is an open benchmark aiming to bridge the gap between the capabilities of generative AI models on English and those on Indian languages.

Google is also developing benchmarks with its partners to assess the quality and performance capabilities of the large language models (LLMs) the company has been building, Gupta said.

Apart from IndicGenBench, the tech giant recently released QuestBench benchmark to determine if LLMs can pinpoint the single, crucial question needed to solve logic, planning, or math problems.

Google DeepMind is also in talks with many partners for creating similar LLM benchmarks for other use cases, Gupta said.

Google Cloud's AI enhancements

Gupta shared that Google Cloud has released more than 3,000 product advances in the last one year and its Infrastructure has expanded to 42 regions, across two million miles of cables.

“We've had over 4 million developers now using Gemini. Last year, we've seen 20X increase in Vertex AI usage driven by these core models such as Gemini,” he said.

Gupta also cited examples of companies like Kraft Heinz that accelerated its campaign creation from weeks to hours using Gemini.

In April, Google also launched its seventh generation tensor-processing unit (TPU) ‘Ironwood’ to speed up AI workloads. The newest TPU can deliver 3,600X better performance and is 15X more energy efficient than earlier TPUs.

It delivers 36.7 exaflops of compute per pod, as compared to the world’s fastest super computer offers 1.7 exaflops of performance.

An exaflop is a measure of performance of a supercomputer that can calculate at least one quintillion floating point operations per second.

“The rest of the world talks about GPUs as the only way for speeding up AI workloads. At least in our team, we use GPUs when we can't get hold of TPUs because TPUs are extremely powerful,” Gupta said.