The timing was deliberate. On February 20, during India's AI Impact Summit in Bengaluru, a little-known startup called Sarvam AI unveiled Indus—a chatbot built atop a 105-billion-parameter multilingual model designed specifically for Indian languages and contexts. Two weeks later, in a move that surprised some observers, the company open-sourced the model weights under Apache 2.0, essentially handing over the technical keys to anyone who wanted them.
It was, perhaps, an admission of sorts. Taking on ChatGPT and Gemini with a closed, proprietary system would require capital and compute resources far beyond what Sarvam has raised. But giving away the model? That creates an entirely different game.
The bet is straightforward, if not simple: that India's linguistic complexity and cultural nuances represent a defensible moat against global AI giants, and that open-source infrastructure could accelerate adoption faster than a subscription-based consumer app ever would. Whether that proves prescient or merely hopeful will likely determine not just Sarvam's fate, but the broader trajectory of what's being called "sovereign AI" across emerging markets.
The Multilingual Problem No One Else Solved
Indus handles text and voice across 22 Indian languages, a technical challenge that goes beyond simple translation. The app lets users type or speak queries and get responses back in the same language—including code-switching, that fluid blending of Hindi, English, and regional dialects that characterizes everyday conversation across much of urban India. It's the kind of pattern that trips up models trained primarily on Western datasets.
Early hands-on testing from Business Today noted the assistant identifying itself as running on the 105B model with a knowledge cutoff of June 2025. Users can upload PDFs and images for analysis, though some beta testers reported missing image upload functionality in early builds, the kind of rough edge you'd expect from a startup racing to market. Document drafting, editing features, and a "web simplified" browsing mode round out the feature set.
Sarvam's internal benchmarks, published in March, claim the model wins 90 percent of pairwise comparisons on Indian language tasks against unspecified competitors. Take that with appropriate skepticism—these are vendor claims, not independent audits. Still, the pitch centers squarely on local accuracy for India-specific queries, the territory where global models often falter when confronted with regional idioms, cultural references, or simply the messy reality of how people actually communicate.
A Mixture of Experts, Literally
Under the hood, Sarvam-105B uses a Mixture-of-Experts architecture with 128 sparse experts and Multi-head Latent Attention. The company trained it on 12 trillion tokens spanning code, mathematics, multilingual text, and web data, with heavy emphasis on the ten most-spoken Indian languages. Context windows run at 128K for the 105B variant and 32K for the smaller 30B model.
Company-reported benchmarks position the model as competitive on reasoning tasks: Math500 at 98.6, AIME-25 at 88.3 (96.7 with tools), and BrowseComp at 49.5. TechCrunch noted that the app sometimes takes longer to respond because it runs reasoning steps in the background, with no user-facing toggle to disable that mode in the initial release. Whether users will tolerate that latency for better accuracy remains an open question.
The technical specifications matter less, perhaps, than what happened on March 6. That's when Sarvam released open weights for both the 30B and 105B models, hosting them on AI Kosh and Hugging Face. The Indian Express framed this as a bet on the viability of open-source AI in a field where closed, proprietary models have captured most consumer and enterprise attention.
Follow the Money (and the Absence of It)
Here's where the strategy starts to make more sense. Sarvam raised $41 million in combined seed and Series A funding back in December 2023, led by Lightspeed with participation from Peak XV and Khosla Ventures. That's the last confirmed funding round in mainstream press as of March 2026—more than a year old, which is an eternity in the current AI fundraising cycle.
Open-sourcing the model weights could be a way to build ecosystem traction and demonstrate technical credibility without requiring the massive capital outlays needed to compete directly with OpenAI or Google on consumer infrastructure at scale. The Apache 2.0 license allows commercial use, which could accelerate adoption among developers building India-focused applications or services that need robust multilingual support.
It's a pragmatic pivot, if that's what it is. You can't out-capital the hyperscalers, but you might be able to out-ecosystem them in a specific geography.
Partnerships That Signal Strategic Intent
Around the launch window, Sarvam announced two notable integrations that hint at where the company sees its actual opportunity. HMD Global, which manufactures Nokia and HMD feature phones, is bringing Sarvam's assistant to feature phones—targeting users without smartphones. The partnership was highlighted during the India AI Impact Summit and positions the technology for populations that still rely on basic handsets, a market segment the Geminis and ChatGPTs of the world have largely ignored.
Bosch Global Software Technologies signed a collaboration agreement on February 20 to co-develop AI-enabled products in automotive, medical devices, and wearables, along with secure large-scale deployments. TechCrunch mentioned these partnerships in coverage of the app launch, though details on specific products or timelines remain scarce—the sort of strategic announcement that could mean a lot or very little, depending on execution.
Company LinkedIn updates in late March referenced a partnership with Razorpay aimed at "voice-first, conversational commerce" and mentioned an early pilot with Swiggy inside Indus. These claims lack independent press confirmation as of late March 2026, which doesn't mean they're not happening, just that they're not yet material enough to warrant outside coverage.
Beta Adoption and the Rough Edges That Come With It
The Economic Times reported more than 10,000 downloads on Google Play by the evening of February 21, a day after sign-ins opened. That's a point-in-time metric, not a sustained trend, but it suggests some initial curiosity among early adopters. The app is free in beta; no consumer subscription plan has been publicly announced, which raises questions about the eventual monetization path.
Users reported several limitations that speak to the product's early stage. You can't delete chat history without deleting your entire account. There's no way to toggle reasoning mode off, which can slow down responses. The app initially launched in India with sign-in via phone number, Google, Microsoft, or Apple ID. An App Store update on March 5 added dark mode and expanded sign-in access outside India, followed by stability fixes on March 19—the usual cadence of a beta release ironing out problems in public.
Privacy claims in the App Store listing emphasize that infrastructure is built and hosted in India, with data staying local—a feature that matters in a market increasingly sensitive to data sovereignty. The app is developed by Axonwise Private Limited, the corporate entity behind Sarvam. LinkedIn shows the company at 51 to 200 employees, with 309 profiles associated as of late March 2026, though these figures fluctuate and are self-reported, so treat them as directional at best.
The Sovereign AI Thesis
Sarvam's founders bring relevant pedigree to the mission. Dr. Vivek Raghavan worked on population-scale technology at UIDAI, India's biometric identity system that now covers over a billion people. Dr. Pratyush Kumar, an AI researcher who led open-source Indic language efforts via AI4Bharat, complements that with technical depth in multilingual NLP. They launched the company in August 2023 with a mission centered explicitly on sovereign AI for India—locally developed and deployed models that reflect regional languages, cultural context, and data residency concerns.
The term "sovereign AI" has become something of a rallying cry, shorthand for technological self-determination in an era where a handful of American companies dominate the AI infrastructure stack. Coverage across Business Standard, TechCrunch, and Indian Express frames Indus as India's answer to ChatGPT and Gemini, with multilingual support and local tuning as the primary differentiators.
Whether that's enough to carve out meaningful market share in a category dominated by extraordinarily well-funded global players remains the central question. The open-source release suggests Sarvam is betting on developer adoption and ecosystem effects rather than trying to compete head-to-head on the consumer front, which may be the only realistic path forward.
Competition is heating up on multiple fronts. It's not just ChatGPT and Gemini anymore; it's a growing field of regional AI labs and enterprise players all racing to serve the Indian market with varying degrees of localization. Sarvam's move to release model weights under a permissive license could give it leverage among developers and enterprises looking for alternatives to proprietary stacks.
But the durability of that advantage depends entirely on whether the open models can keep pace with rapid advancements from closed competitors who are pouring billions into compute and talent. Open-sourcing is a strategy, not a guarantee. And in AI, the distance between a good model today and an obsolete one six months from now can be uncomfortably short.
