India’s growing artificial intelligence ecosystem received a major boost as Sarvam AI unveiled two new large language models (LLMs) — Sarvam-30B and Sarvam-105B — at the India AI Impact Summit on 18 February. The announcement marks a significant milestone in the country’s push toward sovereign AI infrastructure and homegrown foundation models.

Positioned as open-source alternatives to foreign AI systems, the two models are designed to deliver high-quality performance while improving efficiency and reducing token usage — a critical factor in lowering operational costs for enterprises and developers.

Strengthening India’s Sovereign AI Ambitions

The launch comes at a time when India is accelerating its efforts to build domestic AI capabilities under the government-backed IndiaAI Mission. With growing concerns around data security, infrastructure costs, and reliance on overseas technology providers, initiatives like Sarvam AI’s are seen as strategic steps toward technological self-reliance.

By releasing both models as open source, Sarvam AI aims to foster adoption among:

Developers building AI-powered applications

Enterprises seeking localized AI solutions

Government agencies prioritizing data sovereignty

Research institutions advancing AI innovation

This approach encourages transparency, collaboration, and customization — elements that are crucial for scaling AI within regulated and multilingual markets like India.

Built from Scratch: A Fully Indigenous Approach

According to Pratyush Kumar, Founder and CEO of Sarvam AI, the models were developed entirely from the ground up.

He emphasized that the systems were created without external dataset dependencies, underscoring the company’s focus on developing indigenous AI capabilities. This foundational independence aligns with broader national goals of reducing technological reliance on foreign ecosystems.

Sarvam-30B: Efficiency-Focused Large Language Model

The Sarvam-30B model is the smaller of the two but has been engineered with efficiency at its core.

Key Highlights:

Pre-trained on 16 trillion tokens

Supports a 32,000-token context window

Designed to generate high-quality responses using fewer tokens

Token efficiency is a critical metric in large language models because it directly impacts compute requirements and operational costs. By optimizing response quality while reducing token consumption, Sarvam-30B aims to make large-scale AI deployment more cost-effective for enterprises.

During the summit, Sarvam AI presented benchmark results indicating that Sarvam-30B outperformed several well-known AI systems in specific evaluation categories.

Sarvam-105B: Advanced Model for Long-Context Intelligence

The larger Sarvam-105B model significantly expands capabilities with a 128,000-token context window.

What This Means:

Ability to process longer documents

Handle extended multi-turn conversations

Support enterprise-scale knowledge management tasks

According to the company, Sarvam-105B performs on par with other frontier models of comparable size — including both open-source and proprietary systems. The expanded context window makes it particularly suitable for sectors such as legal services, research, governance, and enterprise documentation.

Real-World Demonstration: AI on a Feature Phone

One of the most compelling moments at the launch was a live demonstration showcasing a chatbot named “Vikram” running on a basic feature phone equipped with a physical keypad.

The demonstration highlighted:

Multi-language conversational capability

Operation in low-resource hardware environments

Practical applicability beyond high-end smartphones

The chatbot interacted in Hindi and Punjabi, reflecting the company’s focus on supporting India’s linguistic diversity.

The name “Vikram” pays tribute to Vikram Sarabhai, the renowned Indian physicist widely regarded as the father of India’s space programme. The naming underscores the broader theme of technological self-reliance and national innovation.

Open Source Strategy: Driving Adoption and Innovation

By making Sarvam-30B and Sarvam-105B open source, Sarvam AI is positioning itself within a growing global movement toward transparent AI development.

Open-source models provide:

Greater customization for enterprise needs

Improved auditability and compliance

Faster innovation through community contributions

Reduced vendor lock-in

For India, this could translate into stronger AI ecosystems tailored to local languages, governance requirements, and business environments.

Implications for India’s AI Ecosystem

Sarvam AI joins a growing group of domestic startups working to build large-scale AI models optimized for India’s multilingual and enterprise landscape.

The launch signals several broader trends:

Increased focus on AI sovereignty

Investment in domestic compute and infrastructure

Expansion of open-source AI frameworks

Development of AI systems tailored to emerging markets

As India aims to position itself as a global AI hub, the introduction of high-parameter, open-source foundation models marks an important step in that direction.

The Road Ahead

The unveiling of Sarvam-30B and Sarvam-105B reflects a strategic shift toward building AI systems that are not only powerful but also efficient, accessible, and locally governed.

If widely adopted, these models could accelerate AI innovation across sectors ranging from education and governance to finance and telecommunications. More importantly, they reinforce India’s ambition to shape its own AI future rather than relying exclusively on global technology giants.

With open-source accessibility, extended context capabilities, and a focus on token efficiency, Sarvam AI’s latest models represent a significant contribution to India’s evolving artificial intelligence landscape.