India has officially entered the global foundation AI race.
Bengaluru-based startup Sarvam AI has launched two large open-source models Sarvam-105B and Sarvam-30B at the AI Impact Summit. These models are built from scratch, trained on trillions of tokens, and specially designed for India’s languages, infrastructure, and real-world use cases.
This launch is being called a major moment for India’s “Sovereign AI” ambition the idea that India should build and control its own AI systems instead of depending on foreign companies.
The Two New Models: 105B and 30B
Sarvam AI has introduced two powerful models, each designed for different purposes.
1️⃣ Sarvam-105B – The Brain
- 105 billion parameters
- 128,000-token context window
- Built for complex reasoning, mathematics, coding, and enterprise use
- Can understand very long documents and hold long conversations
Sarvam claims that this model competes with global leaders like Google’s Gemini and Meta’s Llama models in reasoning and accuracy and in some benchmarks, even performs better.
This is Sarvam’s flagship model built for heavy thinking and advanced tasks.
2️⃣ Sarvam-30B – The Speedster
- 30 billion parameters
- 32,000-token context window
- Trained on 16 trillion tokens
- Optimized for real-time, cost-efficient applications
This model is built for situations where speed and lower computing cost matter more than extreme reasoning power.
Sarvam says it outperforms models like Google’s Gemma 27B and Mistral models in multilingual tasks — especially for Indian languages.
Why “Sovereign AI” Is So Important
Today, many companies in India rely on APIs from companies like OpenAI or Anthropic.
That means:
- Data may leave the country.
- Critical systems depend on foreign servers.
- Policy changes from other countries can affect local businesses.
Sarvam’s co-founder Vivek Raghavan highlighted that AI sovereignty is more important than just building the biggest model.
With support from the government’s IndiaAI Mission, Sarvam has built its stack inside India. This means:
- Indian data stays in India
- Infrastructure runs domestically
- Sensitive sectors like defense, banking, and governance get better privacy protection
This is a major step toward technological independence.
Built for India: Where Global Models Struggle
Unlike general-purpose models like GPT-4o, Sarvam’s models focus deeply on Indian needs.
🌐 1. Strong in Indian Languages
The models support:
- 22 Indian languages
- Code-mixed speech like Hinglish (Hindi + English)
Most global models struggle with Indian language diversity and informal speech. Sarvam is built specifically to handle this.
🤖 2. Agentic Capabilities
These models are not just chatbots.
They are designed to act as AI agents that can:
- Plan travel itineraries
- Execute multi-step coding tasks
- Perform structured workflows
- Handle enterprise automation
This makes them useful for real-world productivity tasks.
🎤 3. Voice-First Design
In India, many users prefer voice over typing.
Sarvam models are optimized for:
- Low-latency voice interaction
- Real-time commands
- Accessibility for non-technical users
This makes AI more inclusive for rural and first-time internet users.
Full-Stack AI Ecosystem
Sarvam is not just releasing models and stopping there. They are building a full ecosystem.
🖥️ Domestic Compute Infrastructure
Sarvam is partnering with Yotta to ensure that the hardware infrastructure runs inside India.
This strengthens the Sovereign AI vision.
📄 Sarvam Vision
Sarvam also introduced Sarvam Vision, a document-understanding model.
It reportedly beats Gemini and GPT-4 in OCR (Optical Character Recognition) for complex Indian documents — which is important for:
- Government paperwork
- Banking documents
- Legal files
- Regional language forms
📱 Built for the “Last Citizen”
One of the most impressive demos at the summit showed an AI agent running on a basic feature phone with a physical keypad.
This shows Sarvam’s focus is not just urban smartphone users — but also people in rural areas who still use simple mobile phones.
Their goal is to make AI accessible to every Indian citizen.
Can Sarvam Compete Globally?
Sarvam AI is currently valued at around $200 million. That is small compared to trillion-dollar giants like Google and Microsoft.
In pure general-purpose power, it may not yet beat the biggest global models.
But Sarvam is not trying to win the entire world.
It is trying to win India.
By building:
- High-performance models
- Cost-effective solutions
- Privacy-compliant systems
- Deep support for Indian languages
Sarvam has shifted India’s AI story from being just a user of global AI tools to becoming a creator of foundational AI technology.
Leave a Reply