INDUS by SARVAM AI

Indus is a multilingual conversational artificial intelligence (AI) assistant and chatbot application developed by the Indian startup Sarvam AI. Launched in limited beta on 20 February 2026, Indus serves as The Primary consumer interface for Sarvam AI's flagship large language model (LLM), Sarvam 105B, with support for text and voice interactions. Designed primarily for users in India, Indus emphasizes support for multiple Indian languages, code-mixing, voice-first interactions, and cultural relevance. It is positioned as a "sovereign" AI product, built and hosted entirely in India to prioritize data sovereignty and local optimization.

History

Indus was launched on 20 February 2026, shortly after Sarvam AI announced its Sarvam 105B and Sarvam 30B models at the India AI Impact Summit in New Delhi (February 2026). The app provides a chat-based interface to these models, initially in limited beta with a waitlist due to compute constraints. Sarvam AI, founded in 2023 in Bengaluru, develops foundational AI models tailored to Indian languages and contexts as part of national efforts toward sovereign AI capabilities.

Features

Indus supports text and voice inputs, with responses in text and audio (NATURAL speech synthesis). Key features include:

• Native support for over 22 Indian languages, including seamless code-switching (e.g., mixing Hindi and English).
• Voice-first interactions and document analysis (uploading and interpreting files or images for text extraction).
• Optimization for Indian contexts, including cultural nuance and accessibility on varied devices.

As of March 2026, Indus remains in early beta with limitations such as restricted chat history deletion and occasional wait times.

Mobile apps

Indus is available as mobile apps on iOS and Android, launched in beta in February 2026. Users sign in via phone number, Google, Microsoft, or Apple ID (initially limited to India). The apps support voice input/output and mirror the web interface.

Agents

Indus itself focuses on conversational chat. Sarvam AI's related product, Samvaad, enables deployment of conversational AI agents for tasks like appointment booking, payment follow-ups, and CRM integration across voice, WhatsApp, and web. These agentic capabilities are enterprise-oriented and not yet fully integrated into the consumer Indus app as of March 2026. No advanced autonomous agents (e.g., browser control or multi-step tasks) have been announced for Indus.

Images

As of March 2026, Indus does not include native image generation. It supports document/image analysis for text extraction and interpretation, but no generation of new images.

Training

Indus is powered by Sarvam AI's in-house models:

• Sarvam 105B: A 105-billion-parameter mixture-of-experts model trained from scratch in India, optimized for multilingual tasks, reasoning, and efficiency.
• Sarvam 30B: A lighter variant for faster responses.

Both models were open-sourced under Apache 2.0 in March 2026 and use a custom tokenizer for Indic scripts.

Infrastructure

Indus runs on Indian sovereign compute infrastructure, emphasizing data localization and enterprise-grade security (e.g., private cloud VPC or on-premises options for regulated use). Unlike global services with extensive international GPU clusters, Sarvam focuses on efficient, localized deployment with limited initial capacity leading to beta waitlists.

Languages

Indus supports over 22 Indian languages with native proficiency, including handling of code-mixed conversations common in India. It is designed for linguistic diversity, differing from global chatbots optimized primarily for English.

Limitations

As an early beta product powered by models smaller than many global frontier systems, Indus has several documented limitations. Users cannot delete individual chat histories (full deletion requires account removal), and The Reasoning mode cannot be disabled, potentially adding latency to simple queries. Like other LLMs, Indus is subject to inherent probabilistic limitations, including potential hallucinations (nonsense or inaccurate outputs presented as fact). Sarvam addresses this through grounding mechanisms and focus on Indian-context alignment, though residual risks remain. No widespread reports of successful jailbreaking, major security incidents, or watermarking controversies have emerged for Indus as of March 2026, though general LLM vulnerabilities apply.

Age restrictions

Indus is designated for users aged 13 and above, with no built-in parental consent attestation or stricter under-18 blocks reported as of March 2026. This contrasts with some global competitors that have faced scrutiny over impacts on minors.

Model versions

Indus is powered by Sarvam AI's current flagship models (as of March 2026). No extensive historical version table exists yet, unlike more mature products.

Both models were trained from scratch on Indic-optimized datasets and released under Apache 2.0.

Reception

Indus received significant attention following its launch at the India AI Impact Summit 2026, with Sarvam's pavilion drawing record crowds and media describing it as a milestone for India's sovereign AI ambitions. Early coverage praised its multilingual capabilities, cultural alignment, and potential to reduce reliance on foreign AI, positioning it as a homegrown alternative to ChatGPT and Gemini amid India's growing AI adoption (ChatGPT has over 100 million weekly users in India). Downloads reached 10,000–50,000 in the first weeks, with positive feedback on Indic language handling. Cofounder Vivek Raghavan noted challenges in competing with free global tools, emphasizing the need for India-specific value. No major artistic, political misuse, or regulatory controversies (e.g., bans or influence operations) have been reported for Indus as of March 2026, unlike some global counterparts.

See also

• Sarvam AI
• Large language model
• Artificial intelligence in India