Articles

At the ongoing India AI Impact Summit in New Delhi, Bengaluru-based startup Sarvam AI introduced two advanced large language models (LLMs) developed specifically for India’s multilingual ecosystem. The announcement signals a strong push towards building sovereign AI systems that reflect the country’s linguistic diversity and digital needs.

Two Scalable Models for Diverse Applications

Sarvam AI launched:

• A 30-billion-parameter model, engineered for real-time conversational use. With a 32,000-token context window, it enables extended dialogues while keeping inference costs manageable. The model is positioned for long interactions, enterprise workflows, and agent-based applications.

• A more powerful 105-billion-parameter model, supporting up to 128,000 tokens. This expanded context allows it to undertake complex reasoning, multi-step problem-solving, and long-form analytical tasks.

Both systems are built using a mixture-of-experts architecture, a design that selectively activates portions of the model during computation. This approach significantly reduces operational expenses without compromising performance.

According to the company, the larger model demonstrates competitive results on benchmarks assessing mathematics, coding proficiency, and structured problem-solving when compared with international models such as Gemma, Mistral, and Qwen. Despite being considerably smaller than the 600-billion-parameter DeepSeek R1, the 105B model was trained from the ground up and delivers comparable levels of reasoning capability. The company also indicated that it offers a cost advantage over Gemini Flash, while outperforming it on selected evaluations.

Designed for India’s Linguistic Reality

A defining strength of Sarvam’s models is their full support for all 22 scheduled Indian languages. They have been trained on trillions of tokens drawn from diverse and high-quality datasets, including literature, financial records, newspapers, archival materials, and mixed-language content such as Hinglish.

Recognising that a large segment of Indian users prefer speaking over typing, the models are optimised for voice-first interactions, enabling intuitive access across linguistic backgrounds.

In addition to text generation, Sarvam AI has developed strong visual intelligence capabilities. Its multimodal model can perform image captioning, scene text recognition, chart interpretation, and detailed table parsing. The system reportedly achieved over 84% accuracy in document intelligence tasks involving Indian scripts, demonstrating its potential in digitisation and data extraction across sectors.

Vikram: A Multilingual Conversational Assistant

At the Summit, the company also presented Vikram, a multilingual chatbot capable of seamless conversations across Indian languages. The name pays tribute to eminent physicist Vikram Sarabhai, reflecting the spirit of indigenous scientific advancement.

The development and training of these models were supported under the government-backed IndiaAI Mission, with infrastructure contributions from Yotta and hardware support from Nvidia. This collaboration highlights the growing synergy between public initiatives and private innovation in building India’s AI infrastructure.