Sarvam And The Sovereign AI Dream

In March 2024, the IndiaAI Mission was launched with an outlay of INR 10,372 Cr to build a domestic AI ecosystem. The timing was ripe, but the ambitions were audacious. 

At the time, global leaders such as OpenAI, Anthropic, Google, Mistral, and Meta were already deep into the AI arms race, releasing increasingly powerful models and tools, armed with tens of billions of dollars in investment. 

Think OpenAI, which had already raised more than $18 Bn by October 2024, while Anthropic and Mistral had secured multi-billion-dollar backing.

Against this backdrop, India’s allocation seems modest. The INR 10,372 Cr ($1.1 Bn) was meant to fund not just model development, but also data infrastructure, GPU access, and research labs. Spread across these priorities, the capital looked insufficient to compete with firms that had individually raised more than India’s entire AI budget. 

Undeterred, the flagship IndiaAI Mission continued to tread its path. The mission first lined up hardware (GPUs), and then selected 12 organisations and consortia to build multiple indigenous foundational models. This signalled a distributed, ecosystem-driven approach rather than a single national champion. 

Critics questioned whether this would deliver globally competitive results. Two years later, the script has flipped in India’s favour, and writing are Indian AI startups. 

India’s LLM Contestant Arrives

At the recently concluded India AI Impact Summit, Sarvam AI emerged as the most visible outcome of the IndiaAI Mission. 

Founded in August 2023 by IIT alumni Vivek Raghavan and Pratyush Kumar, Sarvam hit the national headlines when it raised a nifty $41 Mn within five months of its incorporation. But despite early funding and a couple of product launches, Sarvam struggled initially to build momentum and faced scepticism over its pace, ambition, and lower downloads. 

However, the perception changed in 2025, when it was selected to undertake the ambitious project of building an Indian large language model (LLM). And the startup finally delivered on a big note at the AI summit. 

Ahead of the event, the company executed an aggressive “14 days, 14 launches” campaign, emulating OpenAI’s rapid release strategy in late 2024. The campaign culminated with the unveiling of two foundational LLMs: Sarvam-30B and Sarvam-105B.

These marked India’s formal entry into the foundational LLM race.

Sarvam-30B is designed as a lightweight, cost-efficient model. It supports a context length of up to 32,000 tokens and has been trained on nearly 16 Tn tokens. The company positions it as an efficient reasoning and coding model, optimised for lower consumption. 

Benchmark tests indicate that Sarvam-30B performs competitively with global rivals in its size class, including models from Google, Mistral, Alibaba, and Nvidia. In several reasoning and coding tasks, it matches or approaches their performance. 

On the other hand, Sarvam-105B, the startup’s large model, supports a 1,28,000 token context window, enabling more complex, agent-like reasoning. According to internal evaluations, it performs on par with several frontier, open source and closed models in its category.

Technically, these are credible achievements as training models at this scale, with limited capital and domestic infrastructure, reflect strong engineering execution. 

Does India Now Have A Global LLM?

In niche areas, particularly Indic languages, Sarvam’s models demonstrate impressive performance. They often outperform global models, such as GPT-4o, Gemini 3 and Llama 70B, in Indian language tasks, efficiency benchmarks, and localised reasoning. 

However, this does not yet translate into global leadership. 

Frontier models from OpenAI, Anthropic, Google, and Meta operate on a different scale altogether. Many are trained on trillions of tokens, exceed one trillion effective parameters through mixture-of-experts architectures, and are refined through years of interactive deployment. As a result, they dominate benchmarks such as MMLU, GPQA, and complex agentic evaluations. 

While Sarvam is competitive, it typically trails in raw reasoning depth, long-context reliability, and hallucination resistance beyond the Indic domain.

However, the startup’s “India-first” strategy prioritises sovereignty, efficiency, and localisation. This makes its models highly relevant for domestic governance, enterprises, and vernacular applications. But it also means sacrificing some universality and global generalisation. 

In an interview, Raghavan acknowledged this trade-off, calling the launches a ‘significant first step’ and admitting that building models at the scale of Gemini or Claude requires far more capital. 

Sarvam Goes All In

The Indian AI startup’s ambitions extend beyond AI models. At the summit, it unveiled Kaze, AI-powered smart glasses that can listen, interpret, respond, and capture visual context. The company frames this as a shift from screen-based AI to wearable, ambient intelligence. 

If successful, this could position Sarvam as a full-stack AI company rather than a pure model developer. 

The summit also highlighted that India’s progress is not limited to Sarvam. 

Gnani.ai launched Vachana TTS, a voice cloning system that supports 12 Indian languages with as little as 10 seconds of audio input. Meanwhile, the IIT Bombay-led BharatGen consortium unveiled Param2, a 17 Bn mixture-of-experts multilingual model. Both of them were also part of the IndiaAI Mission.

Together, these efforts indicate a growing, diversified AI research base. 

What distinguishes India’s approach is cost efficiency. Rather than matching Silicon Valley’s spending, India attempted to build sovereign models optimised for governance, public services, and domestic enterprises. 

Despite the progress, structural challenges remain formidable for these emerging LLM builders. Also, global AI giants are investing tens of billions annually, acquiring startups, building more advanced LLMs, and shipping new plugins. In comparison, Indian startups operate with constrained capital, limited commute access, and a thinner global distribution. 

For companies like Sarvam, Gnani, and BharatGen, the path forward will require sustained public and private funding, long-term access to compute, talent retention within India, a stronger industry-academia interface, and open deployment to build developer ecosystems. Without these, India risks producing capable but regionally confined models. The question is: do we really want to restrain our flight?

 

Edited By Shishir Parasher
Creatives: Abhyam Gusai

The post Sarvam And The Sovereign AI Dream appeared first on Inc42 Media.