Sarvam AI Unveils Groundbreaking Open-Source AI Models Built Entirely in India
In a significant development for India's technology landscape, Bengaluru-based artificial intelligence startup Sarvam AI has released two powerful AI models as open-source. The Sarvam 30B and Sarvam 105B models are now available for anyone to download, use, and build upon completely free of charge, marking a major milestone in indigenous AI development.
Launch and Recognition at India AI Impact Summit
These models were officially launched at the prestigious India AI Impact Summit held in New Delhi last month, where they received substantial praise from global technology leaders. Notably, Google CEO Sundar Pichai commended the achievement, highlighting the growing capabilities of India's AI ecosystem. What makes these models particularly remarkable is their complete indigenous development – every aspect from conception to deployment was handled within India using domestic resources.
The models represent a comprehensive achievement in domestic AI development, having been built entirely from scratch in India without relying on foreign frameworks or adaptations. Sarvam AI managed the entire process independently, including data collection, model architecture design, training methodologies, and final deployment strategies. The computing power was provided under the government's IndiaAI Mission infrastructure, while the massive training datasets were meticulously curated in-house by Sarvam's research teams.
Co-founder Pratyush Kumar emphasized the significance of this accomplishment, stating that it serves as definitive proof that "India can build state-of-the-art AI models from scratch" without depending on foreign technology or expertise. This achievement positions India among a select group of nations capable of developing advanced AI systems independently.
Dual Model Architecture Serving Different Purposes
Sarvam AI has developed two distinct models optimized for different applications and use cases:
- Sarvam 30B – The Speed Specialist: Designed for rapid response and real-time conversational applications, this model operates similarly to Google's Gemini Flash in terms of efficiency. It currently powers Samvaad, Sarvam's conversational agent platform, and can process up to 32,000 tokens of text simultaneously – approximately equivalent to the length of a short novel. This makes it ideal for customer service applications, quick information retrieval, and interactive dialogue systems.
- Sarvam 105B – The Heavy Lifter: Engineered for complex, multi-step reasoning tasks, this model functions comparably to Gemini Pro in capability. It supports an impressive context window of 128,000 tokens and serves as the foundation for Indus, Sarvam's flagship AI assistant designed for sophisticated problem-solving. Both models utilize an innovative mixture-of-experts architecture that activates only specific portions of the model's parameters during operation, maintaining high performance while significantly reducing computational costs.
Benchmark Performance Against Global Competitors
Sarvam's models demonstrate exceptional performance metrics that place them among world-leading AI systems:
- Mathematical Reasoning: Sarvam 105B achieved an outstanding score of 96.7 on the challenging AIME 2025 mathematics competition test, demonstrating superior quantitative reasoning capabilities.
- Coding Proficiency: On the LiveCodeBench v6 programming benchmark, the model scored 71.7, establishing its competence in software development tasks.
- Agentic Task Performance: For real-world action tasks like web browsing and software tool integration, Sarvam 105B scored 68.3 on the Tau2 benchmark, surpassing OpenAI's GPT model which achieved 65.8.
- Academic Excellence: In a particularly impressive demonstration for Indian educational contexts, the model scored 70 out of 75 on JEE Mains 2026 in a single attempt and achieved a perfect score of 75 in two attempts.
The company has further claimed that Sarvam 105B outperformed both Google's Gemini 2.5 Flash and DeepSeek R1 – a massive 600-billion parameter Chinese model – on specialized Indian language technical benchmarks, highlighting its regional optimization.
Designed Specifically for Indian Users and Languages
What truly distinguishes Sarvam's models from international competitors is their deep integration with Indian linguistic and cultural contexts. Both models comprehensively support all 22 official Indian languages and are specifically optimized for voice-first interactions, recognizing the growing importance of voice interfaces in the Indian market.
The training process incorporated trillions of tokens of Indian-language data, including substantial amounts of Hinglish – the blended Hindi-English language commonly used in everyday conversations, messaging applications, and digital communications across India. This linguistic focus ensures the models understand regional nuances, colloquial expressions, and context-specific meanings that global models often miss.
This development represents more than just technological achievement – it signifies India's growing self-reliance in critical technology sectors and provides domestic developers, researchers, and businesses with powerful tools tailored to local needs. The open-source nature of these models ensures they can be freely adapted, improved, and integrated into various applications across education, healthcare, governance, and enterprise solutions.
