“A lot can happen in six months,” Ola chief Bhavish Aggarwal, revealed a bunch of announcements about its AI venture Krutrim at Sankalp 2024. The launch of Krutrim six months ago has led to several chains of events, including the Microsoft Azure Cloud and Google Maps exit. Now, Krutrim is finally showing its love for Indian developers.
Apart from a bunch of announcements about electric bikes and the launch of BharatCell, the event was fully dedicated to Krutrim’s announcements and future roadmap.
According to the company, since its launch six months back, Krutrim Cloud has received around 250 billion API calls and has around 25k developers using the platform, with one trillion tokens generated so far. Aggarwal also added that the company aims to build India’s first AI cloud in its own data centres which will be powered by the company’s own Bodhi chips, slated to be released by 2026, with Bodhi 2 by 2028.
Moreover, Krutrim Cloud will be free for developers till Diwali this year.
Krutrim has announced a significant expansion of its cloud services, unveiling over 50 new offerings on its Krutrim Cloud platform. These services are designed to cater specifically to the needs of Indian developers and include AI Pods, AI Studio, a language hub with multimodal translation capabilities, and customer experience AI.
Full Stack AI Platform
Back in May, the AI unicorn launched its first LLM, Krutrim-7B-chat, trained on ten Indian languages, on the Databricks Marketplace for free. This was announced by Gautam Bhargava, VP and head of AI engineering at Krutrim.
This time, Bhargava announced that Krutrim has the best Indic tokeniser when compared to GPT-4o and Gemini, but falls behind when compared to Llama 3.1.
Moreover, speaking about the speech models, Krutrim model is on par with AI4Bharat’s IndicTrans2, but beats Gemini and Azure. When it comes to visual world understanding, Krutrim’s model competes closely with Idefics-2.
Krutrim is also launching another AI developer app called Bhashik, a multimodal language hub that can translate videos into different languages. “This event is currently being live-streamed in multiple languages through the Bhashik platform,” said Aggarwal. He also said that edtech startup Unacademy is also leveraging Ola Krutrim’s translation API for its language learning app.
During a scripted demo, Ola Krutrim previewed an upcoming image recognition feature that will be added to the AI model later this year. Aggarwal also showcased how Dashtoon, an early stage startup, uses Krutrim AI Cloud for training and fine-tuning its models.
When it comes to Ola Maps, Aggarwal mentioned that it is expanding its API for developers to encompass over 95 per cent of use cases, including routing, places, maps, tiles, and SDKs. Moreover, another startup, Together, is using Ola Maps for enhancing its customer experience, competing with MapmyIndia. “It is the best maps out there,” added Aggarwal.
For developers, Ola unveiled AI Pods for affordable GPU access, an AI Studio for creating complex AI applications, a model catalogue featuring LLMs and vision models, as well as capabilities for no-code/low-code training, fine-tuning, inference, and model evaluation.
Starting today, Krutrim Cloud on AI Studio will be available for developers and provide no-code and low-code computing platforms tailored for Indian use cases, along with offerings like GPU-as-a-service and model-as-a-service, with Krutrim H100 tiny for 1/25th of the price on AI Pods, which is an abstraction service with GPU slicing for resource efficiency.
Moreover, Ola has also launched the Udaan startup programme. Startups will get Ola Maps credit worth up to Rs 50 lakhs free for three years on Krutrim Cloud and access to mentorship and Ola’s partner ecosystem.
The India Chip Revolution Begins
When it comes to building an AI chip in India, Sambit Sahu came on stage and said that the company is on the roadmap to challenge SOTA AI chip performance by 2028 with Krutrim Silicon. Sahu unveiled three distinct chip families—Bodhi for AI, Sarv for general computing, and Ojas for edge computing.
The Bodhi chip is specifically engineered for complex AI workloads. It focuses on enhancing the speed and efficiency of AI systems. By 2028, Krutrim aims to launch Bodhi 2, an advanced version expected to rank among the top-performing AI chips worldwide. It will be able to run ten trillion parameter models.
Krutrim also plans to produce its first AI silicon chip by 2026.
To achieve its chip development goals, Krutrim has formed strategic partnerships with leading technology firms Arm and Untether AI. “Basically, we would be the best in class in performance per rupee and performance per watt,” added Aggarwal.
When comparing image classification capabilities, Sahu showcased a demo where Krutrim’s Bodhi 1 was able to classify one million images faster than the leading AI chip, possibly NVIDIA’s offering. Adding to that, Aggarwal said that Krutrim will set up a gigawatt scale data centre by 2028. Today, the company’s capacity is at 20 megawatt.
“There is no need for developers to look outside Krutrim,” said Aggarwal.
Ironically, it looks like Ola Krutrim is also trying hard for the sovereign AI approach, similar to Sarvam AI, and naming its products strikingly similar to already existing ones in the market like Sarv-1 with Sarvam and Bhashik with Bhashini.
(FYI: AIM was present at Sankalp 2024 representing the developer community)
