Sarvam AI launches made-in-India foundational AI models, takes first step towards competing with ChatGPT

India's Sarvam AI launches its first homegrown foundational models, offering enterprise-grade AI with support for Indian languages. Here are all the details.

Bengaluru-based artificial intelligence startup Sarvam AI has launched two large language models, Sarvam-30B and Sarvam-105B, as it seeks to establish India’s presence in the global AI landscape. The announcement, made on Wednesday at the AI Impact Summit 2026, marks a notable step for Indian startups aiming to develop domestic AI infrastructure that can compete with established systems such as ChatGPT and other international models. The models are designed for reasoning, multilingual communication, coding, research, and enterprise applications.

Foundational AI models are large-scale neural networks trained on extensive datasets to understand and generate human-like text. Once trained, these models can be adapted for specific applications including coding support, chatbots, content generation, translation, and long-form summarisation. By creating such models in India, Sarvam AI addresses key concerns for domestic enterprises: controlling costs, ensuring data privacy, and supporting multiple local languages, without relying entirely on global APIs. These models can serve as the backbone for various AI-driven services across sectors such as finance, governance, research, and education.

It is worth noting that GPT models, like OpenAI’s GPT-4, are also foundational AI models, trained on massive datasets to perform a wide range of tasks from text generation to reasoning. Sarvam’s models follow the same principle but are designed with Indian languages, enterprise use cases, and accessibility in mind.

Sarvam-30B is the smaller of the two models, with 30 billion parameters and a context window of 32,000 tokens. It has been trained on 16 trillion tokens and designed to operate efficiently, delivering high-quality responses while using fewer computational resources. Benchmark results shared at the launch show the model performing competitively against international models including Gemma 27B, Mistral-32-24B, Nemotron-30B, Qwen-30B, and GPT-OSS-20B across tasks measuring mathematical reasoning, coding accuracy, and general problem-solving.

The model is intended for widespread accessibility. In a live demonstration, Sarvam showcased “Vikram,” a multilingual AI chatbot capable of conversing in Hindi, Punjabi, Marathi, and other Indian languages, even on feature phones. The demonstration highlighted how AI could reach users beyond smartphones, enabling conversational and content generation services for populations previously unable to access advanced AI applications.

The Sarvam-105B model, with 105 billion parameters and a 128,000-token context window, is designed for more complex reasoning and large-scale analytical tasks. During the summit, the model analysed a company balance sheet in real time, providing precise answers to detailed financial and contextual questions. Its ability to process both structured and unstructured data simultaneously positions the model for enterprise use cases such as analytics, research, coding assistance, and summarisation of long documents. This level of performance demonstrates that Indian-built models can match international systems in handling sophisticated business tasks.