Sarvam-105B

Learning Objectives

• Understand what Sarvam-105B is and why sovereign AI models matter
• Evaluate the model's multilingual capabilities across India's 22 official languages
• Assess the role of government support in building nation-specific AI infrastructure

What Is Sarvam-105B?

Sarvam-105B is India's first fully domestically trained 100+ billion parameter open-source language model. Released in beta on February 20, 2026, at the AI Impact Summit in New Delhi, it supports all 22 official Indian languages — from Hindi and Bengali to Tamil, Telugu, Kannada, and Malayalam.

Built by Sarvam AI (a Bengaluru-based startup founded by former AI4Bharat researchers at IIT Madras), the model was trained on 4,086 NVIDIA H100 GPUs provided through India's government-backed IndiaAI Mission — making it a prime example of how government investment can bootstrap a nation's AI capabilities.

Architecture and Specifications

The MoE architecture means Sarvam-105B activates only about 10 billion parameters per query — giving it the knowledge breadth of a 105 billion parameter model at the inference cost of a much smaller one.

Why It Matters

India has 1.4 billion people speaking 22 official languages (and hundreds of dialects). Until Sarvam-105B, no AI model adequately covered this linguistic diversity:

• The US frontier flagships (OpenAI, Anthropic) — optimized primarily for English; limited Indian language support
• Llama and Mistral — open-source but trained predominantly on English and European language data
• Sarvam-105B — trained with substantial allocation to the 10 most-spoken Indian languages, with coverage of all 22 scheduled languages

This makes Sarvam-105B essential for Indian government services, education, healthcare, and commerce — where citizens interact in their native languages, not English.

Government Support

Sarvam AI was selected as one of 12 organizations under India's IndiaAI Mission:

• 4,086 NVIDIA H100 GPUs provided for a 6-month training period through public-private partnership
• 246.72 crore rupees (~$29 million) in government compute and financial support
• Open-source weights published on India's national AIKosh platform for other organizations to build upon

Company Details

Strengths

• Only model covering all 22 Indian languages — no Western or Chinese model matches this linguistic breadth for India
• Open source (Apache 2.0) — freely available for Indian startups, government agencies, and enterprises to build upon
• MoE efficiency — 105 billion parameters with only ~10 billion active per query; cost-effective inference
• Government backing — IndiaAI Mission support validates the model for public sector use
• Potential unicorn — reported $1.5 billion valuation round with NVIDIA would make it India's first AI unicorn

Limitations and Considerations

• Beta status — released in beta (February 2026); final production release timeline unclear
• Funding still pending — the $250 million round at $1.5 billion is reported but not confirmed as closed
• Benchmark comparisons limited — Indian language AI benchmarks are less standardized than English ones; direct comparison to frontier models is difficult
• Small team — startup with ~$41 million raised (pre-pending round); limited capacity versus well-funded competitors
• Government GPU access is temporary — the 4,086 H100s were provided for a 6-month period; long-term compute strategy unclear

Key Takeaways

• Sarvam-105B is India's first sovereign 100+ billion parameter AI model — open-source, supporting all 22 official Indian languages, trained on government-provided NVIDIA GPUs
• MoE architecture (105 billion total, ~10 billion active) provides frontier-class knowledge at efficient inference costs
• Critical for serving India's 1.4 billion people in their native languages across government, education, healthcare, and commerce
• In talks to raise $250 million from NVIDIA at a $1.5 billion valuation — would be India's first AI unicorn