There’s a specific frustration that most Indian developers know. You’re building something for Hindi users, or Tamil users, or literally anyone who isn’t typing in English. You use one of the big Western AI models. And the thing just… doesn’t work right. Not because the model is bad. But because the model wasn’t built with your users in mind. Not even close.
This is basically where Sarvam AI started. Not from a business plan. From two researchers who had spent years measuring exactly how badly global AI models failed for India, and who got tired of measuring.
So they decided to build something instead.
The People Who Started It (And Why They Were the Right Ones)
Vivek Raghavan and Pratyush Kumar are not typical startup founders. Both are from IIT backgrounds. Raghavan did his PhD at Carnegie Mellon and Kumar did his at ETH Zurich. Between the two of them, you have decades of work on language AI, digital infrastructure, and research at places like IBM, Microsoft, and IIT Madras.
Raghavan’s story is kind of unusual for someone who ended up as an AI founder. After his PhD, he joined a company doing chip design automation. Then around 2011, he made a decision that made very little financial sense. He walked into UIDAI and started helping build Aadhaar’s biometric systems. For free. He stayed there for about twelve years. If you know anything about Aadhaar and how complicated it was to get a billion people enrolled in a digital identity system, you understand what that actually means. He was building population-scale infrastructure before that phrase became a startup buzzword.
Kumar’s path was different. He spent years at IBM Research and Microsoft Research, and between 2018 and 2021 he was faculty at IIT Madras. While at IIT Madras, he got deeply involved with AI4Bharat, a research lab working on open-source AI for Indian languages. This is where the two of them ended up working closely together.
And this is actually where the core problem became very clear to both of them.
The Token Tax Problem (This Is the Whole Thing)
There’s a concept called “tokenization” in AI. Basically, models don’t read words the way humans do. They break text into small pieces called tokens. The problem is that English text gets broken into tokens very efficiently. One English word is roughly 1.4 tokens on average.
For Hindi or Kannada, the same word takes 4 to 8 tokens to process. This is what people in the field started calling the “token tax.” And this isn’t just a weird technical detail. It has real consequences. If your language costs 5x more tokens to process, the AI runs slower and costs more per call. For a startup building for Indian users, or a government trying to deploy AI at scale, that’s not a small difference.
AI4Bharat had spent years measuring this gap. Their benchmark IndicXTREME covered 20 Indian languages across nine tasks. The result was damning: Indian languages performed three to four times worse than English on standard evaluations. Everyone working in this space knew the problem. But nobody well-funded was actually fixing it from the ground up. There were fine-tuning projects, there were API wrappers, there were startups doing niche applications.Nobody was training base models from scratch on Indian data.
By 2023, four things had changed: hundreds of millions of Indian users were now on the internet through vernacular apps; Aadhaar and UPI had already proven that population-scale digital infrastructure was possible; voice was emerging as the natural interface for people who didn’t read or type easily; and large language models had made language the main way humans interact with software.
Raghavan and Kumar looked at this and decided to leave what they were doing. In August 2023, they co-founded Sarvam AI, which is a Sanskrit word meaning “everything” or “all.” The name makes sense when you understand what they were trying to do.
The First Year: Funding and the First Attempt
In December 2023, Sarvam raised approximately $41 million in a combined seed and Series A round, led by Lightspeed Venture Partners, with Peak XV Partners and Khosla Ventures also coming in. For an Indian AI startup at the time, this was a big number. But honestly, in the context of what it actually costs to train frontier AI models, $41 million is not a lot. Most US labs are spending more than that on a single training run.
So the team had to be smart about it.
Their first significant model, Sarvam-M, released in May 2025, was built on top of Mistral Small, a French AI company’s base architecture, and fine-tuned with Indian datasets. It was a pragmatic decision. And it worked well enough for the products they were shipping. But the critics were not wrong. If the base model is French and you’re just fine-tuning on top, how sovereign is it really? That’s a fair question.
I think the team knew this too. Because what came next was much more ambitious.
February 2026: The Models That Changed the Conversation
On February 18, 2026, at the India AI Impact Summit in New Delhi, Sarvam unveiled two foundational large language models trained from scratch on Indian datasets and Indian compute infrastructure: Sarvam-30B and Sarvam-105B.
This is where things get interesting. Let me explain the difference.
Both models were trained from scratch. Not fine-tuned from someone else’s base. Pre-training, supervised fine-tuning, reinforcement learning, all of it done internally. Training was conducted entirely in India on compute provided under the IndiaAI Mission.
Sarvam-105B uses a Mixture-of-Experts architecture. It has 105 billion total parameters but only about 10.3 billion are active for any given token. This design is the same general approach that made DeepSeek so disruptive when China released it. You get a large model’s capacity but at a fraction of the inference cost, because most of the model is “sleeping” at any given moment. It’s clever architecture. And it’s especially important for India, where you need cheap inference if you want AI to actually reach a billion people.
Sarvam-30B, the smaller one, powers Samvaad, their conversational agent platform. The 105B powers Indus, their AI assistant built for complex reasoning and agentic workflows.
The 105B’s performance on some benchmarks was genuinely surprising to a lot of people. It scored 98.6 on Math500 and wins 90% of pairwise comparisons in Indian language evaluations. On the JEE Main January 2026 paper, the 105B got a perfect score on the mathematics section. That’s not trivial, JEE is one of the hardest engineering entrance exams in the world.
That said, I should be honest about the limitations here. The Hacker News thread after the release included reports of hallucination issues and noted that the model’s knowledge cutoff is June 2025, so it has no awareness of events in the second half of 2025 or 2026. For any use case that needs current information, that’s a real gap. Also, on BenchLM’s general leaderboard, Sarvam 105B ranks around 87th out of 124 models, so let’s not pretend it’s competing with GPT-4o on general tasks. It’s not. But for Indian-language reasoning and agentic tasks, it’s doing things that no other model was doing before.
Both models were released open-source under Apache 2.0, which means anyone can download, use, or modify them commercially. This part matters a lot. India gets the infrastructure, not just the product.
What They’ve Actually Built (Beyond the Models)
The models are the headline, but Sarvam was shipping real products well before the 105B came out. This is the part that I think people outside India miss.
Their speech models are transcribing over 500,000 hours of audio every month. Their conversational platform handles over 2 million interactions per day. Their inference platform processes 10 million API calls daily, and usage tripled in just three months.
Sarvam Vision, their document AI model, is being used to digitize over 35 million pages of insurance forms, land records, and other physical documents. If you’ve ever tried to work with a government land record in India, you know how difficult it is. These are often handwritten, sometimes decades old, in regional scripts. Getting AI to read them accurately is not a small achievement.
The agriculture ministry deployment is worth mentioning separately. Multilingual voice agents collected data from 17 million farmers, giving the Ministry of Agriculture and Farmer’s Welfare high-quality feedback that would have been impossible to gather at this scale otherwise. These are farmers who may not read well, may not own a computer, and who definitely aren’t going to type in English. Voice AI in their local language is the only interface that actually works.
One insurance company ran a nationwide voice campaign using Sarvam’s platform to support policy renewals for 45 million policyholders.
So when people talk about AI sovereignty, this is what it actually looks like in practice. Not just models on a leaderboard. Real deployments, at a scale that most countries can’t do.
The Series B and What Just Happened This Month
On June 15, 2026, two weeks ago from when I’m writing this, Sarvam announced the first close of their Series B.
They raised $234 million at a post-money valuation of $1.5 billion. HCLTech came in as the lead strategic investor with $150 million, which also gave them over 10% of the company. Bessemer Venture Partners joined in, with existing investors Khosla Ventures and Peak XV Partners continuing their support. The target for the full round is $300 million.
That makes Sarvam India’s newest AI unicorn.
To understand why HCLTech’s involvement is interesting: HCLTech has deep enterprise relationships globally. The idea is to combine Sarvam’s AI models with HCLTech’s enterprise reach, engineering depth, and software assets to build AI products for businesses and governments outside India too. This is not just a financial investment. HCLTech gets access to Sarvam’s technology. Sarvam gets distribution at scale.
And now the Indian government is also getting into the picture. Reports from today indicate the Centre may end up with a 1–2% stake in Sarvam through the IndiaAI Mission. The arrangement is linked to subsidized GPU compute that the government provided to Sarvam earlier, being converted into equity through compulsorily convertible debentures. This is a new structure for India’s public-private AI collaboration, and people are still debating whether grant-based funding would have been better than equity-based support. The argument against it is that it dilutes founders. The argument for it is that it aligns government incentives with the startup’s success. Honestly, I don’t know which side is right.
The funding also came at a specific political moment. Days before the announcement, the US government had ordered Anthropic to suspend access to some of its advanced models for foreign nationals, citing national security. This put a spotlight on how much India depends on a small number of overseas AI providers. The timing was not lost on anyone.
How Did a Small Team Do All This
This part still kind of surprises me. As of mid-2025, Sarvam had around 114 employees. Even if they’ve grown since then, we’re talking about a company with maybe 200–350 people, depending on which source you check, that has trained two frontier-class models from scratch, built multiple production AI products, deployed them at population scale, and just became a unicorn.
Compare that to the teams at Google DeepMind or Anthropic. Those are organizations with thousands of people working just on AI research.
Part of the reason Sarvam can do more with less is the compute access through the IndiaAI Mission. Part of it is that the founders have unusual backgrounds for AI startup people. Raghavan spent over a decade building Aadhaar. He knows what it means to deploy at population scale without many resources. That’s a different mindset than a startup that’s just trying to get to a B round.
And the research team has real depth. These aren’t people who started doing AI after ChatGPT launched. Pratyush Kumar was publishing on Indian-language AI at IIT Madras years before it was a hot topic.
The training setup itself required serious optimization work. Beyond the datasets, they had to optimize tokenization, model architecture, execution kernels, scheduling, and inference systems to make deployment work across a wide range of hardware, from high-end GPUs to devices like laptops. That’s the kind of systems work that only happens when your team has really internalized the constraints you’re working under.
What Comes Next
The new funding is going toward three things: a next-generation frontier model for agentic, coding, and cybersecurity use cases; coding AI specifically for Indian software engineers and enterprises; and expanding compute access to scale up their deployments.
The coding AI direction is interesting. India has tens of millions of developers. A lot of them are working in codebases that Western coding tools don’t handle particularly well, or in contexts that require understanding of Indian development environments and workflows. Whether Sarvam can carve out a meaningful space here against GitHub Copilot and Cursor is something I genuinely don’t know. That’s a very competitive market.
The agentic direction is where the bigger bets are. Sarvam’s agentic platform already powers a sales enablement tool for a large fintech company that has a 350,000-person sales force. If you’ve ever tried to build an agent that works reliably in Hindi or Tamil across unreliable connections, you know how hard this is. Most Western agentic frameworks just don’t account for this.
The cybersecurity focus is newer and less clear. I’m guessing this is partly about government contracts, where data sovereignty and security are table-stakes requirements.
There’s also a hardware angle developing. At the February 2026 summit, Prime Minister Modi was spotted testing Sarvam Kaze, which are India’s first AI-powered smart glasses. The glasses handle real-time translation across 22 languages. Whether that product actually goes anywhere is a separate question, but it shows Sarvam is thinking beyond the API business.
The India AI story has a lot of hype. Some of it is justified. Some of it isn’t. Sarvam has done something real. Building from scratch, with a small team, on limited compute, and getting to a model that actually competes on reasoning benchmarks while handling Indian languages better than anyone else, that’s not nothing. The bigger test is whether they can sustain that pace now that they have much more money and much more to lose. Usually when companies raise $234 million, they also raise their overhead, their politics, and their coordination costs.
Whether Sarvam keeps the research velocity they had when they were smaller is the thing I’d actually watch over the next 12 months.
For now though, it’s hard to argue with what they’ve built.