After DeepSeek, Krutrim goes ‘open-source’ with new set of AI models

“Our focus is on developing AI for India – to make AI better on Indian languages, data scarcity, cultural context etc,” the Ola co-founder said in a post on X on Tuesday, February 4.

Aggarwal further unveiled plans of building India’s largest supercomputer in partnership with AI chip giant Nvidia by the end of 2025. The supercomputer will be powered by Nvidia’s flagship GB200 chips.

He also announced Krutrim AI Labs to carry out frontier AI research with a focus on “ developing India’s first frontier-scale AI models, creating state-of-the-art multimodal AI systems, and publishing breakthrough research in top-tier AI conferences like NeurIPS, ICML, ACL, Interspeech, and CVPR,” as per the official website.

Announcing the @Krutrim AI lab today! https://t.co/s5DmiGhQoo

While we’ve been working on AI for a year, today we’re releasing our work to the open source community and also publishing a bunch of technical reports.
Our focus is on developing AI for India – to make AI better on… pic.twitter.com/M5nBtlE5p0

— Bhavish Aggarwal (@bhash) February 4, 2025

“While we’ve been working on AI for a year, today we’re releasing our work to the open source community and also publishing a bunch of technical reports,” Aggarwal said. “By open sourcing our models, we hope the entire Indian AI community collaborates to create a world-class Indian AI ecosystem,” he added.

Krutrim’s new AI models

The Krutrim AI lab has released Krutrim-2, an LLM with a parameter count of 12 billion. The company claimed that Krutrim-2 has shown strong performance in processing Indian languages. Parameters are numerical values that are used to indicate the size of an LLM.

As per the company, Krutrim-2 scored 0.95 in accuracy on the existing Indic benchmarks – IndicXTREME, IndicGenBench and IN-22 . It achieved a score of 80% when evaluated on an English language code generation benchmark test called HumanEval Coding.

With an underlying Mistral-Nemo model architecture, Krutrim-2 has been pre-trained on ” a curated mix of English, Indic, code, math, books, and synthetic data,” the company said. It supports a context window of 128K tokens. The context window of an AI model indicates the number of tokens that can be processed at once.

“We followed a multi-stage training procedure, varying the data-mix, context size and batch size at every stage, leading to a stable and efficient model training. After pre-training, the model underwent supervised training for cross-task instruction following and direct preference optimization for alignment,” Krutrim said.

Besides Krutrim-2, the company has also developed a vision language model called Chitrarth 1. This model is powered by Krutrim-1, a 7-billion parameter system that was launched in January last year.

In addition, Krutrim said it is open-sourcing AI models designed for speech translations (Dhwani 1) and text-to-text translations (Krutrim Translate 1) as well as an Indic LLM for use cases like Search and Retrieval-Augmented Generation (RAG) called Vyakhyarth 1. RAG is a machine learning technique used to improve the accuracy of an AI model.

Krutrim AI Labs has also developed a global benchmark called BharatBench for evaluating LLMs on their Indic performance. “We’re nowhere close to global benchmarks yet but have made good progress in one year,” wrote Aggarwal.

The company’s AI announcements come just days after Chinese AI startup DeepSeek claimed that it had achieved a breakthrough in computer efficiency with its reasoning model R1, sending shockwaves across the tech industry.