Sarvam Indus and HMD: What Local Language AI Actually Takes

The HMD Vibe 2 5G launched in India preloaded with Sarvam's Indus AI chatbot — a partnership TechCrunch reported on that was first announced at India's AI summit in February 2026. The phone costs ₹10,999 (~$114), runs Android, and ships with Indus visible on first boot. That's the news.

What's worth thinking about is what the partnership actually reveals: what it costs technically to build a genuine local-language AI product, and what the distribution math looks like for a market where AI adoption is still at the start.

What Sarvam actually built

Indus runs on a 105-billion-parameter model trained on Indic language data. That scale puts it in the same range as large commercial models — not a lightweight mobile classifier. The investment makes sense once you understand the actual problem.

The technically significant feature is mid-sentence code-switching. A user might start a question in Hindi, switch to English for a technical term, and finish in Hindi. That is not edge-case behavior; it is how people actually speak across South Asia. A model that cannot track those switches without losing context is practically unusable for the target user.

The same pattern applies directly to Sri Lanka. Sinhala-English mixing is standard in any professional or urban setting. A product that forces users to pick one language and stay in it treats normal speech as a bug.

Training at this scale requires serious infrastructure. Sarvam is reportedly raising $300 million at a $1.5 billion valuation. That is not a detail to gloss over — it calibrates what it actually costs to solve the full problem from scratch.

Why pre-loading on a $114 phone is the strategy

HMD held about 4% of India's feature phone market in 2025 but has close to no smartphone share. This partnership is partly about survival: a software differentiation story in a hardware market where brand recognition is thin.

For Sarvam, the math is different. Indus has around 293,000 downloads in India. ChatGPT has 43.9 million. The gap is not primarily a quality gap — it is a discovery gap. Users who might benefit most from a local-language assistant are the least likely to seek out a new app on their own. Pre-loading bypasses that entirely.

HMD CEO Ravi Kunwar put it plainly: get the app in front of consumers by having it there on first boot. Distribution beats discovery in early adoption, especially when the target market is not yet looking for what you built.

One honest limitation worth flagging: Indus currently does not work offline and has no device integration shortcuts. For users in areas with patchy connectivity — which is part of the stated target — that is a real gap. It is the kind of thing that gets addressed in later versions, but version one is incomplete by its own terms.

What this means if you are building for Sri Lanka

Sri Lanka has two official languages with different scripts, plus English woven through both. The code-switching challenge is at least as complex as what Sarvam is solving for Hindi-English. The difference is resources.

Sarvam has institutional backing that no Sri Lankan AI team currently has. But "we cannot build a 105B-parameter model" does not mean "we cannot build useful local-language tools." The constraint changes the approach.

Fine-tuning an open-source base model — Llama 3, Mistral, or similar — on Sinhala and Tamil data is tractable for a small team today. You are not starting from zero: both languages are represented in multilingual training datasets, and the fine-tuning infrastructure (LoRA, QLoRA on a single consumer GPU) is well documented. If you want a baseline for what Sinhala text processing looks like in a browser right now, the image-to-text OCR tool at induwara.lk extracts Sinhala and Tamil text from scanned documents using Tesseract LSTM, entirely client-side and without a server call. It is a narrow utility, but it shows the building blocks exist.

The result of a fine-tuned model will not be a general assistant. It could be a useful domain-specific tool: a Sinhala-first customer support agent, document Q&A for government forms, or a voice interface for a specific vertical. Tractable problems with a real audience.

The hardware context transfers too. The Vibe 2 5G at $114 is roughly the price point that dominates mid-range Android in Sri Lanka. If you are designing an AI feature for a Sri Lankan audience, the device is a low-end Android on LTE. Anything that assumes a flagship chip or fast fixed-line connection will not reach the users who need local-language support most.

What this means for you

The HMD-Sarvam story is useful not as inspiration but as a spec sheet. Local AI at scale requires a large model trained on real local-language data, a distribution channel that does not rely on users finding you, and patience with the gap between early adoption and where the global incumbents already are.

The first of those is hard to replicate without serious funding. The second — distribution — is something any builder targeting Sri Lanka has to think about from day one. The third is just where things are: 293,000 downloads versus 43.9 million is not failure; it is what early looks like.

If you are working on anything in the Sinhala or Tamil NLP space, Sarvam's approach — large multilingual base model, code-switching support, domain-specific deployment — is now a public blueprint. The components exist. The remaining question is the right vertical and the right way to make the product reach the people it is supposed to help.