Ola’s Krutrim AI launches ‘open-source’ model to take on DeepSeek, to work closely with NVIDIA

Ola’s Bhavish Aggarwal-led Krutrim AI is making waves with the release of new open-source AI models. India’s growing ambition to be a strong contender in the global AI race, currently ruled by the US and China, got a major boost with this announcement. Aggarwal revealed plans to invest more than $230 million into the startup, with a goal to secure an additional $1.15 billion in funding by next year.

Aggarwal also emphasised Krutrim AI’s mission to create AI tailored to India’s needs, addressing challenges like language diversity, limited data, and cultural nuances. The company also aims to build the nation’s largest supercomputer by 2025 in collaboration with NVIDIA, leveraging the chip giant’s top-tier GB200 processors.

Krutrim AI’s open-source vision

The release of Krutrim AI’s latest models was described as a call to action for the Indian AI community. Aggarwal shared his excitement about open-sourcing their work to encourage innovation and collaboration. He highlighted the launch of Krutrim AI Labs, which will focus on cutting-edge research, including large-scale AI models and multimodal systems that integrate multiple forms of data such as text, images, and speech.

Krutrim AI Labs has already rolled out its latest language model, Krutrim-2, which boasts 12 billion parameters. According to the company, the model excels in handling Indian languages, achieving a near-perfect accuracy score on benchmarks like IndicXTREME and IN-22. Krutrim-2 also performed well on a global coding test, scoring 80 per cent in generating code based on human instructions.

More from Tech

Amazon, Microsoft, Alphabet, and Meta to spend well over $300 billion on AI in 2025, fearing China China’s DeepSeek, Trump are playing spoilsport at the Paris AI Action Summit

Next-gen AI models with a local focus

Krutrim-2 is based on a Mistral-Nemo architecture and has been trained on a diverse blend of data, including English and Indic languages, mathematics, and synthetic material. The company explained that a multi-stage training process was used to ensure stable and efficient model development. The AI model can process up to 128,000 tokens in a single session, making it capable of complex, large-scale tasks.

Additionally, Krutrim AI has introduced several other models to diversify its offerings. The Chitrarth 1 vision-language model builds on the capabilities of Krutrim-1, which launched last year with 7 billion parameters. For speech and text-based tasks, Dhwani 1 and Krutrim Translate 1 have been open-sourced, along with Vyakhyarth 1, an Indic language model designed to enhance search and retrieval tasks using advanced machine learning techniques like Retrieval-Augmented Generation (RAG).

A push for Indian AI excellence

To measure how well AI models perform in Indian contexts, Krutrim AI has developed a new benchmark called BharatBench. Aggarwal acknowledged that while Krutrim has made promising strides within a year, there is still progress to be made to compete with global standards.

This announcement comes just after Chinese AI startup DeepSeek unveiled a breakthrough model in computational reasoning, raising the stakes in the global AI industry. As India accelerates its AI development efforts, Krutrim AI’s latest initiatives mark a significant step towards establishing a stronger foothold in the tech world.

Editor’s Picks

1

India to launch its own foundational AI model in few months, over 18000 GPUs deployed: Minister Ashwini Vaishnaw