At its Zoholics 2025 event, Zoho made two significant announcements for the Indian market, its in-house large language model (LLM) dubbed Zia LLM and a bilingual Automatic Speech Recognition (ASR) system called Zia ASR. Both models are positioned as enterprise-ready AI solutions built ground-up by Zoho's India teams.
Speaking at the event, company executives revealed that the Zia ASR will support speech-to-text in both English and Hindi, a clear nod to India’s diverse linguistic landscape and the rising need for localised AI solutions.
The introduction of Zia LLM, which includes models scaled up to 7 billion parameters, marks Zoho’s foray into building foundational language AI in-house. “This isn’t just an integration of third-party tech, we’ve built these models ourselves for accuracy, data security, and real-world enterprise use,” a senior Zoho executive stressed.
Additionally, the company showcased a slide indicating its 51% compound annual growth rate (CAGR) in India over the last decade. “India is not just a market, it’s a growth engine,” one of the presenters said, adding that Zoho’s strategy continues to centre on deep R&D investments and regional product relevance.
The launch of these AI models comes amid a broader global trend of companies shifting from off-the-shelf generative AI to more controlled, custom-trained models. In that context, Zoho’s insistence on building and owning its AI stack offers a clear differentiation in the crowded SaaS space.
“This is not about chasing the AI hype,” said another company official. “It’s about building practical, scalable tools that address the real needs of our users, especially in emerging markets like India.”
Looking ahead, company insiders hinted that regional language support may expand beyond Hindi, with early-stage work already underway. This aligns with Zoho’s larger commitment to digital inclusivity and its long-standing philosophy of building for Bharat.
With growing scrutiny around AI data privacy, Zoho also reiterated its policy of zero monetisation of customer data. “Our models train on publicly available or permissioned data only. Customer data is never used to train or fine-tune our models,” the spokesperson pointed out.
