Sarvam AI officially announced the launch of its first foundational large language models, Sarvam 30B and Sarvam 105B, during the India AI Impact Summit in New Delhi.
These models represent a major milestone in India’s “Sovereign AI” push, moving away from being a “wrapper” for foreign models to building high-performance architectures from scratch using local compute.
The New Model Lineup
Both models utilize a Mixture-of-Experts (MoE) architecture, which allows them to be massive in scale while remaining incredibly efficient for real-time use.
| Model | Parameters | Activated Parameters | Context Window | Best For |
| Sarvam 30B | 30 Billion | 1 Billion | 32,000 Tokens | Real-time voice, mobile apps, and feature phones. |
| Sarvam 105B | 105 Billion | 9 Billion | 128,000 Tokens | Complex reasoning, document analysis, and coding. |
Performance vs. Global Rivals
The highlight of the launch was Sarvam’s claim that its models outperform several established global leaders, particularly in reasoning and Indian language tasks.
- Beating DeepSeek R1: Sarvam 105B reportedly outperforms DeepSeek R1 (released in late 2025) across several reasoning benchmarks, despite being roughly one-sixth the size of the 600B-parameter Chinese model.
- Surpassing Gemini Flash: Co-founder Pratyush Kumar stated that Sarvam 105B is cheaper and faster than Google’s Gemini 2.5 Flash, while delivering stronger performance on technical benchmarks specifically tailored for Indian languages.
- Outperforming GPT-120B: On MMLU-Pro (a harder version of the classic general intelligence benchmark), Sarvam’s 105B model scored higher than the popular GPT-120B variant.
- 30B Reasoning Champion: Sarvam 30B was shown to outperform Gemma 27B, Mistral-32-24B, and Qwen-30B in mathematical reasoning and coding accuracy.
Key Innovations: Vikram and Enterprise Analysis
Sarvam demonstrated the models’ real-world utility with two high-profile live showcases:
- “Vikram” on Feature Phones: Using Sarvam 30B, the team demonstrated a multilingual chatbot named Vikram (after Vikram Sarabhai) running on a basic Nokia feature phone with a physical keypad. It successfully handled voice conversations in Hindi, Punjabi, and Marathi.
- Financial Intelligence: Sarvam 105B was used to analyze a complex company balance sheet in real time, answering nuanced questions about financial health that typically require human auditing expertise.
- No Data Dependency: Pratyush Kumar emphasized that these models were built entirely from scratch without relying on external or synthetic datasets from foreign models, ensuring a truly “sovereign” data trail.
Accessibility and Open Source
To encourage mass adoption across India’s developer ecosystem, Sarvam confirmed it will open-source both models. This allows Indian enterprises and government agencies to run the models on their own “on-premise” servers, ensuring maximum data privacy and cost control.
“We are not just building for the elite. Our goal is to make AI deliver population-scale impact. Whether you are on a high-end workstation or a ₹2,000 feature phone, Sarvam intelligence should be available to you.” — Pratyush Kumar, Co-founder, Sarvam AI.
