Sarvam AI Unveils Sovereign Language Models for Indian Languages and Voice-First Use Cases

In a significant push toward building a sovereign artificial intelligence stack, Sarvam AI on Wednesday unveiled two large language models—Sarvam-30B and Sarvam-105B—positioning them as India’s answer to systems developed by Google, OpenAI and Anthropic. Backed by the government’s IndiaAI Mission, the models were showcased at the AI Impact Summit as open-source systems customised for Indian languages, cultural nuances and voice-first use cases. The company said the move marks a major step toward creating domestically built AI infrastructure that can operate securely within India’s borders. The headline act was Sarvam-105B, a 105-billion-parameter model designed for complex reasoning and enterprise-grade applications. According to the company, the model competes with, and on some benchmarks surpasses, global peers in areas such as mathematics, reasoning and functional correctness. Sarvam-30B, the smaller and more efficient sibling, is built for real-time applications with a context window of up to 32,000 tokens and has been trained on 16 trillion tokens. Unlike many startups that fine-tune existing global models, Sarvam’s founders emphasised that their systems are built from scratch. “Sovereignty matters much more in AI than building the biggest models,” co-founder Vivek Raghavan said. Co-founder Pratyush Kumar added that the focus is on inclusion and scale: “We need to build for the last citizen.” The models are designed to work seamlessly across multiple Indian languages, including “Hinglish,” and are optimised for voice commands — a crucial feature for population-scale deployment. The broader ambition is to position AI as a digital public good, akin to Aadhaar or UPI , enabling large-scale applications in governance, education and commerce. Sarvam describes its technology as full-stack, spanning compute infrastructure, foundational models and enterprise applications. The systems are also “agentic,” capable of performing multi-step tasks such as coding or planning workflows with minimal human intervention — a feature aimed at accelerating enterprise automation. The company claims its 30B model outperforms competitors such as Gemma 27B, Mistral-32-24B, OLMo 31B, Nemotron-30B and Qwen-30B across multiple benchmarks, while being significantly smaller and more cost-efficient than some global counterparts. It also positions its models as more economical alternatives to offerings such as Gemini Flash. The launch comes at a time when India is seeking to reduce reliance on foreign AI infrastructure and build indigenous capacity amid rising geopolitical scrutiny of AI supply chains. Though its immediate focus remains India, Sarvam says it is building for the broader Global South, where affordability and infrastructure constraints mirror domestic realities. Founded in 2023, the startup has raised about $50 million from investors including Lightspeed Venture Partners and Khosla Ventures, and was last valued at roughly $200 million. Whether Sarvam can meaningfully challenge entrenched global leaders remains to be seen. But with a 105B open-source model and a sovereign-first pitch, India’s AI race has acquired a serious domestic contender.