Zoho Corp plans to make two artificial intelligence-based internal foundational models public by the end of the year, group chief executive officer Shailesh Davey said in an interview with Mint.
The cloud software startup is also working on foundational models to support Indian languages, though it does not have a timeline on when they will be launched, Davey said.
The two foundational models will be based on 7 billion and 13 billion parameters, respectively, Davey, who took over as Zoho Corp group CEO in January after the firm’s longstanding co-founder and CEO Sridhar Vembu stepped down to work as chief scientist on the firm’s AI projects, said.
“These are the foundational models we are working on, which we are incorporating in our products, and we will put it out for the public later this year, maybe over the next two quarters,” Davey told Mint on Monday.
“The foundational model is more about a technology demonstrator, and putting it out there is a sign of confidence,” he added.
For companies, one of the challenges for building foundational models has been the higher demand for computational power as well as training it with better quality data.
Zoho has been working on “right sizing” foundational models so that the power consumption for these models is more efficient.
Unlike B2C foundational models, B2B models work in a constrained manner specific to the business requirements and use cases of a particular sector. Therefore, these models may not require large amount of parameters. For context, ChatGPT-4 had 175 billion parameters, while ChatGPT-5 is expected to have 1.5 trillion parameters, according to media reports. Chinese foundational model DeepSeek R-1 had 671 billion parameters when it was released in January, but it also released a smaller version with 1.5 billion to 70 billion parameters, according to a Tech Crunch report.
Currently, Zoho offers customers access to various models, including DeepSeek and OpenAI, up to 32 billion parameters through its Zoho platform. Here, customers pose a query to the platform which taps the appropriate model for a response.
Later this year, two foundational models that Zoho is working on will also become available to its customers through the Zoho platform, the company said.
“The advantage of running a smaller model is that you don’t need so much GPU (graphics processing unit). You make the model smaller and if it does not impact the accuracy for the customer, you can run it in the CPU itself,” Davey said at the media roundtable on Monday. GPUs are used to perform high-level machine learning tasks and are more powerful than a CPU (central processing unit). GPUs also draw more power than a CPU.
Expanding AI solutions
Zoho is also launching a set of Agentic AI product suites for its information technology clients through its enterprise division ManageEngine, which specifically caters to the IT divisions within enterprises. ManageEngine is expected to earn the firm a billion dollars in revenues by 2025 or 2026, Rajesh Ganesan, CEO, ManageEngine, which is a division of Zoho Corp, said.
ManageEngine brings in around 55-60% of the revenues for Zoho Corp, Davey added. Zoho crossed $1 billion in annual revenues in November 2022, according to an announcement made by co-founder Vembu at the time.
For enterprise clients, one of the big focus areas is security, and this is Zoho’s fastest-growing segment within ManageEngine, Davey said. Security for enterprises comprises managing data on phones, laptops, mobile devices, the servers and other access pathways.
Currently, security comprises around 40% of ManageEngine’s revenues and is expected to rise further as it is growing at an annual 60%, Davey said.
