Sarvam's Indic AI model: Hype, Hope and the Hunt for Tech Sovereignty

Artificial Intelligence startup Sarvam AI, which recently rolled out its large language model (LLM), Sarvam-M, is now at the center of a debate. At stake is the question: Is Sarvam AI's new Indic-focused open-source LLM a landmark step in India's AI ambitions, or is it an overhyped product with limited real-world impact?

While some celebrate the model as a breakthrough in India’s journey towards technological sovereignty, others point to low adoption numbers. They further argue that the hype may not match its actual value.

Just to give some context, in April 2025, the Ministry of Electronics and Information Technology (MeitY) selected Sarvam AI to lead the development of India’s first sovereign LLM under the IndiaAI Mission. A month later, on May 23, the company unveiled Sarvam-M, a model with 24 billion parameters designed to understand and generate content in Indian languages.

What Exactly is Sarvam-M?

Sarvam-M is a homegrown alternative to Western AI models. It is based on Mistral-Small, an open-source model from French company Mistral, but heavily adapted to Indian linguistic and cultural contexts. The aim is to bridge the gap in AI models that largely favor English and Western content. Sarvam-M has been trained to understand and respond in 10 Indian languages, including Hindi, Tamil, Bengali, and Marathi.

Read: Changemakers 2024: How Sarvam is Pioneering a Truly Desi AI

In terms of performance, Sarvam AI claims that its model can hold its own against global giants. Despite having fewer parameters than Meta’s Llama 3.3 (70 billion) or Google’s Gemma 3 (27 billion), Sarvam-M delivers comparable or even better results in benchmarks, especially in multilingual tasks. However, with regard to English and general knowledge questions, it slightly trails behind, being about 1% less accurate than the original base model.

The Tweet That Sparked a Firestorm

What set off the broader debate was a tweet from Deedy Das, a venture capitalist at Menlo Ventures. He observed that despite the media attention and the billion-dollar valuation, Sarvam-M had been downloaded only 23 times within two days of its release. In contrast, he highlighted the success of Dia, an open-source model by two Korean students, which had seen 200,000 downloads in just one month.

Das called it "embarrassing" and said that much of India’s AI effort appeared to be copying global trends rather than solving meaningful local problems.

"Nobody is really demanding a marginally improved 24B Indic language model—that much is clear," he wrote, adding that if a team is going to invest in building models, there should be a compelling reason.

He contrasted this with a project by two Korean college students, who had trained an open-source model named Dia that saw around 200,000 downloads in the previous month, calling the comparison “embarrassing.

Last month, Dia was downloaded over 1.6 lakh times, according to Hugging Face. Dia was created by the Korean founders of Nari Labs and is based on a model with 1.6 billion parameters. It can not only speak the words in the text, but also add emotions, tones, and even sounds like laughter, coughing, or clearing the throat, just like a real human would.

His post found traction on X, formerly Twitter, sparking a flurry of opinions on India’s AI ecosystem and how success should be measured.

Read: Why Microsoft-Backed AI Start-Up 'Builder.ai' Is Going Bankrupt - Read Here

Support and Skepticism

Other voices weighed in. Some users argued that Indians tend to prefer English for their digital lives, which may explain the lack of early adoption. Amit Vyas, an AI entrepreneur, suggested that this reflects a cultural aspiration among Indians to communicate and operate in English.

Raj Kunkolienkar, former co-founder of Stoa, highlighted the importance of product storytelling and outreach. "In 2025, when every other day brings a new AI breakthrough, you can't ship a model and expect the world to care just because it handles Indic languages," he wrote. 

He argued that Sarvam needs to go beyond technical merit and showcase real use-cases that resonate emotionally, such as enabling elderly Indians to access technology or bringing regional poetry to life.

He also pointed to the importance of strong marketing, working with influencers, and making the product feel urgent and essential: "The market of dumbos like me is waiting."

A Familiar Fear: The Koo and Hike Parallel

Some critics worry that Sarvam might follow in the footsteps of earlier India-first tech efforts that failed to sustain interest. Think Koo and Hike.

Both platforms were launched as desi alternatives to Twitter and WhatsApp, but neither could sustain traction. Despite Koo achieving over 80 million downloads, it shut down in 2024 after struggling with engagement and monetization. Hike Messenger shut its chat app in 2021, unable to compete with WhatsApp’s ubiquity.

As Kunkolienkar warned, Sarvam AI risks falling into the same "patriotic alternative" narrative trap unless it focuses on building community and demonstrating everyday value.

Others, like investor Surya Kanegaonkar, lament the loss of platforms like Koo. He argued that their user bases could have served as an early feedback loop for tools like Sarvam-M. Without such networks, testing, feedback, and community-building become far harder.

Adding another layer, the Indian government also launched its own bilingual AI model, Param 1, under the BharatGen initiative. Despite being trained on a massive 5 trillion words from Indian sources and tuned for both Hindi and English, Param 1 has only been downloaded 12 times.

The Case for Patience

Not everyone is convinced that download stats are the right yardstick. Investor Pratyush Chaudhury argued that Sarvam-M represents a significant technical feat under constrained conditions. India doesn’t have easy access to powerful GPUs like H100s, which are widely used in the US and China. With export restrictions on the horizon, Indian startups face even steeper challenges.

Moreover, unlike English-centric datasets like CommonCrawl, Indian language data is extremely scarce. Chaudhury mentioned that less than 0.01% of these public datasets represent Indian languages. In this view, what Sarvam AI achieved is not just another AI release—it is a foundational milestone. Comparing it to viral open-source projects may be misleading, especially when judged merely on short-term download metrics, he added. 

Amid the debate going on, one thing that is understood is that the ongoing debate right now, reflects a broader tensions in India’s tech ecosystem—between building for local relevance and competing on global metrics. It remains to be seen how it takes shape over time.