‘I Was Wrong About Sarvam’

For the past year, Sarvam has been talking about sovereignty. Over the last week, it has been on a shipping spree. The latest release, Bulbul V3, a text-to-speech model for Indic languages, and has been topping all charts on benchmarks and social media. 

It led a blind study with over 20,000 votes, handling 11 languages with natural-sounding voices and low error rates. Around the same time, Sarvam Vision clocked 84.3% accuracy on messy documents across 22 Indic scripts.

Right now, Sarvam is having a moment. There is visible hype, and people are betting big on the startup’s vision of sovereign AI. That makes the contrast with last year impossible to ignore. Back then, the verdict was blunt and public.

“Embarrassing.” That was how Deedy Das, partner at Menlo Ventures, described Sarvam’s flagship launch.

Sarvam had just launched Sarvam-M, a 24 billion parameter model built on top of Mistral’s model. Sarvam had government backing, serious funding and a mere 334 downloads on Hugging Face.

In a world where two college students in Korea could rack up two lakh downloads over a weekend, Sarvam looked like a cautionary tale. Too much hype. Too little demand. Another sovereign AI pitch that failed to survive contact with reality.

This week, the same voice returned with a different sentence.

“I was wrong about Sarvam,” Das said.

In tech, public reversals like this are rare. People usually move on quietly. This one did not. And that is what makes it worth examining.

What Did Sarvam Get Right This Time?

Nothing magical happened. Sarvam did not suddenly win the global LLM race. It stopped trying to win the wrong one. 

“When I wrote about them a year ago, I felt like the direction to train small ‘indic’ language models was wrong. But boy, have they turned it around,” Das wrote on social media.

Instead of chasing English benchmarks, Sarvam went deep into the mess most frontier labs ignore, something it had almost promised early on, but which people noticed only now. That was exactly the point Hemant Mohapatra, partner at Lightspeed India Ventures and an investor of Sarvam, made last year. “Most folks really don’t understand what strategy we are going for and what matters in that strategy,” Mohapatra had written on X. 

“There is absolutely a huge demand for what we are building, but looking at it from the generic lens of downloads isn’t rooted in what really works in India,” he said.

Indian speech. Indian scripts. Bad scans. Noisy calls. Code-mixed Hinglish. Government forms, receipts and ledgers. The unglamorous stuff, or in AI terms, the applied stack.

Adithya S Kolavi, founder of CognitiveLab, has been building independent evaluation tools and leaderboards for Indic models. He tested Sarvam Vision himself.

“I tested the model as well, and it’s doing pretty good. [It’s] very, very hard to get that level or performance with a three-billion model,” he tells AIM.

Text-to-speech with Bulbul. Speech-to-text with Saaras. Vision models and ASR models that read and speak 22 Indian languages. OCR that actually understands tables instead of collapsing them.

The shift shows up in the details, not the marketing. That doesn’t mean last year’s criticism lacked merit. It clearly helped the company move in the right direction.

Sarvam Vision, built on Sarvam’s three-billion-parameter model trained from scratch, is less ‘cool demo’ and more ‘back office automation’. On internal benchmarks, it reports word accuracy north of 90% across major Indic scripts, while staying competitive even on English-heavy documents.

For India, that matters more than yet another chatbot. Big labs rarely care, even when they have been trying to build the models for the Indian context.

Their maths does not justify obsessing over Kannada OCR or Marathi speech synthesis. Indian languages make up just 1-3% of training data. Everything is written in English first and translated later.

Sarvam is building natively, and that’s something the IndiaAI Mission has been pushing with the 12 startups it has selected to build foundational models. The difference is starting to show up in sentiment. BharatGen, Tech Mahindra, and others are also set to release more models this month. 

On X, Abdulmuiz Adeyemo, Founder of Nigeria-based Tradia AI, wrote, “Niche dominance beats general competence—owning Indic language infrastructure is more valuable than being the 47th English LLM wrapper. Indian tech finally built something.”

That line captures the pivot.

Sarvam went from “embarrassing” to “well done” in less than a year, and the reversal came from the same VC who trashed them, as one user on X noted.

It is further evident in the recent collaborations. Sarvam has signed an MoU with the Odisha government to establish a full-stack sovereign AI park in the state. This is the company’s second state-level partnership, following its January partnership with Tamil Nadu.

Long Way to Go

Much of the criticism around Sarvam comes from comparing it to OpenAI or DeepSeek, even though the problem the company is trying to solve is fundamentally different.

Back in May 2025, many believed Sarvam had chased the LLM race and flatlined. In reality, it had always been focused on applied products. This distinction was also highlighted in the Economic Survey 2026, where the government argued that competing on frontier LLMs is unnecessary given how prohibitively expensive the race has become.

The right race, instead, is for use cases. And now, with the extremely cheap pricing, Sarvam appears to have hit the nail on the head. This is where the company has quietly improved.

Sneha Prabhu, Product Manager, Model APIs at Sarvam, and one of the people who helped build the dashboard, explained the philosophy on X: “make it simple” enough for anyone to understand what’s going on but also powerful and intuitive enough for developers to start building.

Not everything needs to feel like a research paper. Sometimes, it just needs to feel like a product. Engineers within the ecosystem also acknowledged the turnaround.

Harveen Singh Chadha from Sarvam highlighted the Vision team’s work, calling it “SOTA on olmOCR” and noting that the model goes beyond OCR into captioning, scene text recognition, and complex table parsing.

Even policymakers are taking a victory lap. Union Minister Ashwini Vaishnaw wrote in a post, “Our sovereign model strategy is delivering results. Even the most critical reviewers are praising the technologically advanced model released by Sarvam as a part of our AI mission.”

Setting aside the government phrasing, the message is clear: the bet looks less naive today.

There is also a deeper lesson here about how the ecosystem judges AI startups. Last year, downloads were treated as the primary metric. They were clean and easy to compare. However, they were also misleading.

Developers do not download an OCR engine for fun. They deploy it inside call centres, banks, and government offices. Success shows up in contracts, not GitHub stars.

The ‘big lab’ models are often too generalised for Indian communication. Sarvam’s moat is tackling the ‘messy middle’ with heavy code mixing, accents, and background noise. That chaos is not an edge case here. It is the default.

Sarvam is far from perfect. Bengali mistranslations still show up, and instruction-following weakens in low-resource languages. The company acknowledges as much.

Even so, the gap versus generic global models now looks large enough to matter.

Raj Dabre, Senior Research Scientist at Google and Adjunct Faculty at IIT Madras and IIT Bombay, summed it up with a mix of patience and optimism. “Doing something new takes time to gain respect. Respect to Deedy Das for reevaluating his assessment. As I once said, Sarvam is going to only one place: the moon!”

It may sound dramatic, but the direction feels right. Sarvam has finally demonstrated the importance of Indic AI models, and the use cases may now finally come through.