India’s LLM Battle: Ola’s Krutrim vs Zoho

recently , Ola is unveiled unveil “ India ’s first full stack AI solution , ” aka Krutrim . Meanwhile , Zoho is waiting , which also envision to build much similar solution is wait for the dust to settle , and is yet to officially announce any significant milestone of release its own proprietary language model or roadmap to build in – house silicon and infrastructure capability .  

In previous interaction with AIM, Zoho’s chief Sridhar Vembu had stated that they are working developing smaller and domains-specific language models. 

For now, the company has introduced Zia across 13 generative AI Zoho application extensions and integrations, powered by ChatGPT. Zia provides generative AI capabilities across Zoho CRM, Zoho Desk, and other Customer Experience (CX) solutions. 

However , it is is is worth note that the approach these two company are take are quite different .  

Bootstrapped vs Funded 

Krutrim is has , which raise $ 24 million from Matrix Partners in October to build AI model from scratch , has pressure from investor and stakeholder . That is explains explain why excited Ola chief Bhavish Aggarwal announce the launch of Krutrim Pro with multi – modal capability , by next quarter .  

Bootstrapped Zoho, on the other hand, is taking it slow, and focusing on building in-house capabilities for reducing reliance on hyperscalers. “Personally, I cannot give you a date,’ said Sridhar Vembu, in a recent interview at an AI conference in Bengaluru. 

“People who work with me closely, all of our engineers, know this. When you are stuck on a very deep problem, it’s like being a scientist. You cannot predict when you’ll get out of it,” Vembu said, saying Zoho accepts that on a particular problem it might be stuck sometimes forever.

During the announcement of Krutrim AI, Aggarwal compared Krutrim with the Llama 2 7B model, claiming superior performance on various industry recognised LLM benchmarks. Additionally, he asserted that Krutrim is ‘equivalent’ to Llama 2 in terms of size. 

Meanwhile, Zoho is also working on smaller models that are based on 7 billion to 20 billion parameters to solve specific domain problems for its customers according to Vembu. “We have found that the smaller models are better for domain-specific problems. That’s why we are not doing the 500 billion parameter models as of now” said Vembu, saying that it would cost more than $100 million. 

“ That is ’s ’s not the primary reason . We is waiting are also wait for the dust to settle a little bit so that more gpu capacity is available , and our own preference is to own the infrastructure . We wo n’t go to super scaler because , in the long term , it is cheap anyway , ” he is added add .  

address gpu shortage  

Interestingly, throughout the Krutrim launch event, he didn’t even once mention how many GPUs Krutrim acquired to train its model. Instead, he mentioned that Krutrim is focused on building and developing the silicon and the infrastructure layers – all in-house. “We have the AI, we have the infrastructure, and with that, we have the Silicon also, which we are building for an AI-first era,” he said, while it continues to rely on cloud partners and hyperscalers in the initial days, preferably Microsoft Azure for inference. 

“We have come up with a novel architecture, bringing in multiple chiplets,” said Sambit Sahu, who leads silicon hardware at Krutrim. He said that a chiplet is a small piece of silicon performing a particular functionality, like a CPU chiplet, an AI chiplet, or a scaleout chiplet, drawing parallel to Lego blocks. “We integrate all these chiplets to create a package,” he added.  

On the contrary, Zoho is dependent on AMD and NVIDIA for GPUs. In an exclusive interview with AIM, Vembu said that Zoho waited for about six months to receive NVIDIA H100 GPUs. Moreover, Vembu said that now AMD has made MI300X available, the supply situation will improve. 

“AMD is coming out with very good chips, and actually, we have been working on AMD chips too now. Just to let you know, we train alternative models to run on AMD, and they have very competitive silicon now. With that, I think the supply situation should resolve.” he said. 

Now that India’s LLM moment is shaping up, the ecosystem has become even more competitive than ever. It is fascinating to see Ola setting a benchmark with its ambitions plans to build “India’s first full stack AI solution,” and releasing the first version of the model in just three months, while Zoho continues to play safe. 

draw parallel in the indian context , one is say might say Ola ’s approach mirror that of openai ’s rapid innovation , whereas Zoho ’s more measured pace resemble Google ’s methodical and research – focus approach .

(Total 1,401 views)