The first thing early users noticed about Indus wasn't what it could do. It was what it asked for: a language preference screen offering Tamil, Hindi, Bengali—twenty-two options before you ever typed a word. For Bengaluru-based Sarvam AI, that moment on February 20, 2026, represented something larger than an app launch. Call it a wager. That India's 1.4 billion people might prefer AI trained on their linguistic reality, not Silicon Valley's.
Whether that preference translates to actual adoption, though? That's the harder question.
The limited beta went live across iOS, Android, and web—simultaneously, which itself suggested ambition. But within hours, users were encountering waitlists. Co-founder Pratyush Kumar didn't dance around it: "We're gradually rolling out Indus on a limited compute capacity." Translation: building infrastructure for India's linguistic sprawl costs real money, even with government-backed GPU access helping shoulder the load.
What 105 Billion Parameters Actually Gets You
Two days before the app dropped, Sarvam had unveiled the engine underneath: Sarvam-105B, a Mixture-of-Experts architecture with 105 billion total parameters. The technical specs sound impressive—support for 22 Indian languages, context windows stretching to 128,000 tokens, enough runway to chew through lengthy contracts or hold conversations that drift between languages mid-stream.
Early testers picked up on that last bit. The app could shift from English queries to Hindi responses without being asked, maintaining thread across linguistic borders. Business Today's initial hands-on flagged familiar territory: document uploads, PDF parsing, writing assistance, all recognizable to anyone who's spent time with ChatGPT. Voice responses came with what Channel News Asia described as "native-like Indian accents," which matters more than it sounds. In a country where English fluency maps directly to class and geography, accent alone carries weight.
Sarvam's internal benchmarks—98.6 on Math500, 88.3 on AIME-25—come with the usual caveat. Vendor-reported figures, not independently validated. CNA's March 4 testing praised Indic language handling but noted gaps in transparency and safety protocols, concerns that lingered even after Sarvam open-sourced the model weights two days later. The knowledge cutoff of June 2025 also raised eyebrows, a detail that ages poorly in fast-moving tech cycles.
The Incumbent Problem
Here's the uncomfortable backdrop: OpenAI claims 100 million weekly users in India already. Anthropic draws nearly 6% of its global usage from the country. When Google's Sundar Pichai publicly lauded Sarvam's local-language work during that same summit week, the endorsement carried a certain irony—Gemini competes in the exact same market Pichai was praising Sarvam for serving.
App metrics told part of the story. The Economic Times reported 10,000+ Google Play downloads by February 21. Third-party trackers suggested that number hit 50,000+ within a week, though Sarvam hasn't released official counts. Access stayed geographically fenced—CNA reporters needed India-registered Play accounts to test the app as late as early March, which either signals careful capacity management or reveals distribution constraints, depending on how charitable you're feeling.
Credibility Meets Constraints
Vivek Raghavan and Pratyush Kumar arrived at this project with unusual credibility in India's tech policy world. Raghavan had worked on digital public infrastructure including the Aadhaar identity system, the world's largest biometric database. Kumar co-founded AI4Bharat at IIT-Madras. Their $41 million raise in December 2023—Lightspeed, Peak XV, Khosla Ventures—looked substantial at the time. No subsequent funding rounds have surfaced publicly as of March 2026, though that could mean many things.
The "sovereign AI" positioning wasn't just marketing speak. Government partnerships materialized throughout early 2026: a 50-megawatt compute hub in Odisha, a 20-megawatt AI research park collaboration with IIT-Madras in Tamil Nadu. Device partnerships followed—HMD for Nokia-branded phones, Bosch for automotive integration, work with Qualcomm. Smart glasses called "Sarvam Kaze" got demoed at the summit, shipments supposedly coming in May.
On March 6, Sarvam open-sourced both the 30-billion and 105-billion parameter models under Apache-2.0 licenses. Weights went up on AI Kosh and Hugging Face. The timing wasn't coincidental—CNA had questioned the transparency gap just two days prior. Open weights build developer goodwill and ecosystem momentum. They don't automatically translate to consumer app adoption, but they help.
The Friction Points
Beta limitations tell their own story about where resources went. Users can't delete individual chat threads, only nuke the entire account to clear history. No way to toggle off a "reasoning" mode that sometimes slows responses. Sign-in demands a phone number, Google/Microsoft account, or Apple ID—no anonymous browsing here.
Business Today caught localization quirks: numbers spoken in English during otherwise Hindi text-to-speech, the kind of edge case that undermines the whole cultural fluency pitch. The underlying speech tech—Bulbul v3 for text-to-speech, Saaras V3 for speech recognition—claims coverage across 11 and 22 Indian languages respectively, though production quality at that breadth remains somewhat theoretical until tested at real scale.
The Switching Cost Question
Perhaps the central challenge isn't technical at all. Sarvam clearly can build a multilingual model—they've done it. The harder problem: convincing users already comfortable with ChatGPT or Gemini that better Hindi support justifies switching apps, learning new interfaces, accepting current feature gaps.
Language might be necessary but not sufficient differentiation. Compute costs matter. Response speed matters. Ecosystem integrations—the plugins, the third-party tools, the workflow automation—matter. Cultural nuance competes with functional convenience, and convenience has network effects.
Indus launched with a thesis: India's AI future shouldn't require English as the universal translation layer. The company built partnerships, secured government backing, released open weights to build credibility. Whether any of that converts to durable market share depends on execution details no benchmark captures. And on whether limited compute capacity scales faster than user patience wears thin.
The 50,000 early adopters represent a start—or maybe just curiosity. What comes next will reveal whether linguistic sovereignty can overcome incumbent advantage, one conversation at a time.
