The download counter ticked past 100,000 sometime in early April. Not exactly viral, but for a six-week-old chat app built by a Bengaluru startup most Indians had never heard of, it represented something: proof of concept, maybe, or at least proof of curiosity.
Sarvam AI launched Indus on February 20, 2026 as India's homegrown answer to ChatGPT and Google's Gemini—a multilingual assistant designed specifically for Indian languages, running on the company's own massive language model. The timing wasn't accidental. India has emerged as perhaps the most contested territory in the global race for AI adoption. Sarvam's bet is that those users might want something built for them from the ground up, not just localized from Silicon Valley.
Whether that bet pays off is another question entirely.
What You Actually Get
Indus works as the consumer face of Sarvam-105B, the company's flagship large language model—a 105-billion-parameter system that handles text and voice queries across 22 Indian languages plus English. Available on web, iOS, and Android, the app leans heavily into code-switching, the linguistic mashup (Hinglish being the obvious example) that defines how millions of Indians actually talk.
On paper, the feature list reads impressively: document analysis, web search, voice input and output. Early app store listings and press coverage suggested capabilities for uploading images and PDFs, drafting documents, even agent-like task automation down the road.
Reality has been bumpier. Some early adopters reported missing image upload functionality. UI bugs cropped up in those first weeks—the expected turbulence of a staged beta rollout, though that doesn't make it less frustrating for users who showed up early. One particularly odd limitation: you can't delete individual chat histories. It's all or nothing—wipe your entire account or live with every conversation cluttering your archive.
The app also includes what Sarvam describes as a "reasoning" mode, designed for complex queries. There's no toggle to turn it off, which means slower responses even when you're asking something simple. Perhaps an oversight, perhaps a deliberate choice to showcase the model's capabilities. Either way, it's the kind of friction that can sour user experience.
A Business Today reviewer noted in late February 2026 that the assistant reported a knowledge cutoff of June 2025. Current enough at launch, though that window will narrow quickly.
The Engine Underneath
The technical architecture tells you something about Sarvam's ambitions. The company published documentation in March describing Sarvam-105B as a mixture-of-experts transformer with Multi-head Latent Attention—industry speak for a sophisticated system designed to route queries efficiently across specialized subnetworks. It was pre-trained on 12 trillion tokens: code, mathematics, multilingual text, web data. The context window extends to 128,000 tokens, useful for digesting longer documents without losing thread.
On March 6, in a move that surprised some observers, Sarvam open-sourced both the 105B and a smaller 30B model under Apache 2.0 license. The weights went up on Hugging Face and AIKosh, India's national AI repository. The company positions the larger model for complex reasoning and agentic workflows that power Indus, while the smaller one handles lighter conversational tasks.
Sarvam's own benchmarks show strong performance on mathematics problems: 98.6 on Math500, 88.3 on AIME 2025 (jumping to 96.7 with tools enabled), 85.8 on HMMT. The company emphasizes optimization for Indian language tasks specifically, though independent third-party validation of those claims remains thin.
The infrastructure came together with government support—India's IndiaAI Mission provided compute access, Yotta supplied infrastructure, Nvidia kicked in technical backing. All reported in March, all part of a broader push to build indigenous AI capability.
Early Returns
By early April 2026, the Android version had logged over 100,000 downloads with a 4.2-out-of-5 rating from more than 6,000 reviews on Google Play. The iOS version showed a higher 4.7 rating, though from a much smaller sample—around 380 reviews. The app store developer is listed as Axonwise Private Limited, operating from Bengaluru.
Updates rolled out rapidly. iOS version history shows releases on February 19-20 for the initial builds, then February 23, then March 5. That March update added dark mode and, tellingly, updated sign-in instructions for users "outside India"—a hint that interest extended beyond the home market, or at least that Sarvam wanted to accommodate it.
User feedback followed predictable patterns. App store reviews requested better chat history management and improved copy-export functions. Reddit threads from February and March painted a mixed picture: praise for local language adaptation, complaints about language errors and interface glitches. Standard beta territory, but worth noting for anyone evaluating whether to invest time in the platform.
The Bigger Picture
Sarvam didn't just launch a chat app. At the India AI Impact Summit in New Delhi in mid-February, the company laid out broader hardware ambitions. A partnership with HMD aims to bring Sarvam's assistant to Nokia feature phones later this year—voice queries for local information and government schemes on devices without smartphone capabilities. A play for the next billion users, in other words, or at least the several hundred million.
On February 20, Bosch Global Software Technologies signed a memorandum of understanding with Sarvam to co-develop secure AI solutions for automotive applications, medical devices, wearables. The company has also demonstrated prototype smart glasses running its models, though commercial availability remains unclear.
In March, Sarvam launched a startup program offering API credits and priority support for early-stage builders. The API pricing page lists both the 105B and 30B chat models as "free per token" with no usage charges—an aggressive positioning move against established players, though one that naturally raises questions about long-term business model sustainability.
Sovereign AI and Its Discontents
Sarvam markets Indus heavily around "sovereign AI" framing—infrastructure and models designed for India's linguistic and cultural context, not retrofitted from Western systems. The app store listing emphasizes India-hosted infrastructure and data locality, though the company's privacy policy allows for international data transfers depending on jurisdiction and legal requirements. A necessary legal hedge, perhaps, but one that complicates the sovereignty pitch.
The startup was founded by Pratyush Kumar and Vivek Raghavan and has raised roughly $41 million across seed and Series A rounds announced in December 2023. No newer funding has been publicly confirmed as of April 2026, which either means the company is operating within its current runway or staying quiet about additional capital.
The competitive landscape is, to put it mildly, daunting. OpenAI, Anthropic, and Google dominate the market with massive distribution advantages, compute resources that dwarf what Sarvam can access, and brand recognition that comes from being first or biggest. Whether Indian language optimization and local infrastructure prove sufficient differentiation remains very much an open question.
One hundred thousand downloads in six weeks suggests initial interest. But initial interest and sustained adoption are different animals entirely. Sarvam launched with constraints—limited compute, waitlist access in the early days, missing features—but also with intent. The thesis: that India's AI users want something built specifically for them, not adapted from elsewhere.
The next few months will test whether intent and execution can close the gap. And whether users care enough about linguistic precision and local context to overlook the polish and power that the incumbents offer.
For now, Indus exists. That alone, in a market this competitive, counts as an achievement. What it becomes next is the harder question.
