Sarvam AI Powering a Made-in-India AI Revolution
Sarvam AI, a Bengaluru-based AI startup, unveiled two large language models (LLMs) at the India AI Impact Summit 2026 — Sarvam-30B and Sarvam-105B — both built and trained from scratch in India
Sarvam AI was selected by MeitY under the IndiaAI Mission's Innovation Centre pillar to develop an indigenous foundational AI model, receiving government support of ₹246.72 crore
The models are designed specifically for Indian languages and public service delivery, supporting advanced reasoning, multilingual tasks, mathematics, and coding
Union Minister Shri Amit Shah stated at the summit that Sarvam AI "exemplifies why the future belongs to India"
PM Modi publicly lauded Sarvam AI alongside other indigenous AI models as proof of India's innovative capability
Sarvam AI's approach is defined as "sovereign AI" — development, deployment, and governance remaining entirely within India, using Indian compute infrastructure
Sarvam AI — Technical Architecture and Capabilities
Sarvam AI is an Indian AI company co-founded by IIT-Madras alumni Vivek Raghavan and Pratyush Kumar. It operates as a full-stack AI company: building its own foundational language models, speech models, and application layer products. Both models announced at the AI Impact Summit use Mixture of Experts (MoE) architecture — a design that activates only a fraction of parameters per inference, enabling large model scale while remaining computationally efficient.
Key Details
- Sarvam-30B: 30 billion parameter model, Mixture of Experts design
- Sarvam-105B: 105 billion parameter model; activates ~9 billion parameters per token; 128,000-token context window
- Architecture: Mixture of Experts (MoE) — enables scale without proportional compute cost increase
- Language support: 22 scheduled Indian languages plus English; focus on Indic linguistic nuances
- Training: From scratch on India-centric datasets (not fine-tuned Western models)
- Use cases: Voice-based citizen interfaces, government document processing, multilingual chatbots, coding assistants
- Speech models: Sarvam ASR (automatic speech recognition) and TTS (text-to-speech) for Indic languages
Sarvam AI's foundational model launch marks India's transition from AI consumer and application builder to AI infrastructure producer — the models are trained on Indian data, in India, for Indian needs, addressing the fundamental dependency on foreign models for sovereign deployment.
IndiaAI Mission — Innovation Centre and Foundational Model Programme
The IndiaAI Mission's Innovation Centre pillar funds the development of indigenous foundational AI models. Following a competitive selection process, MeitY shortlisted 12 teams from across academia, startups, and research institutions. Sarvam AI received ₹246.72 crore in government support — primarily computing credits on the IndiaAI compute pool (38,000+ GPUs across 14 cloud providers). This model of public compute as a subsidy for AI development mirrors France's AI Compute programme and the US NSF National AI Research Resource.
Key Details
- IndiaAI Mission total budget: ₹10,300 crore; approved by CCEA, March 2024
- Innovation Centre: 12 teams selected for foundational model development
- Sarvam AI government support: ₹246.72 crore (compute + development support)
- BharatGen Param2 funding: ₹988.6 crore (largest single allocation)
- GPU infrastructure: 38,000 GPUs via 14 cloud service providers; data centres in Mumbai, Navi Mumbai, Hyderabad, Bengaluru, Noida, Jamnagar
- Nodal agency: Digital India Corporation (DIC) under MeitY
- BHASHINI integration: Sarvam's language models feed into the national BHASHINI platform
Sarvam AI's ₹246.72 crore government allocation enabled it to train large-scale foundational models on domestic compute rather than relying on OpenAI or Google APIs — the IndiaAI Mission's compute pool is the direct material enabler of its "Made in India" claim.
Large Language Models (LLMs) — Foundational Technology Concepts
A Large Language Model (LLM) is a neural network trained on massive text datasets to perform natural language tasks. Foundation models are large, general-purpose models trained at scale that can be fine-tuned for specific applications (question-answering, translation, code generation, document analysis). The Mixture of Experts (MoE) architecture, pioneered at scale by models like Mistral Mixtral and Google Gemini 1.5, uses a "router" to activate only a subset of specialised "expert" sub-networks for each input, dramatically reducing inference compute while retaining large model capacity.
Key Details
- Parameter count and capability: Larger models generally more capable; GPT-4 estimated at 1.8 trillion parameters; Sarvam-105B comparable to mid-range international models
- MoE vs Dense models: Dense models activate all parameters per token; MoE activates only a fraction (Sarvam-105B activates ~9B of 105B per token)
- Context window: Sarvam-105B has 128,000-token window — equivalent to ~90,000 words; enables processing of long documents
- Training data: Quality and diversity of training data (not just scale) determines model quality for specific languages and domains
- Evaluation benchmarks: MMLU (general knowledge), HumanEval (coding), IndicEval (Indian language tasks)
Understanding MoE architecture is key to appreciating why Sarvam's 105B-parameter model is computationally viable on India's current GPU infrastructure — MoE reduces the effective compute cost of inference, making large Indic-language models practically deployable at government scale.
- Company: Sarvam AI, Bengaluru; co-founders include IIT Madras alumni Vivek Raghavan and Pratyush Kumar
- Models unveiled: Sarvam-30B and Sarvam-105B (both MoE architecture, trained from scratch in India)
- Sarvam-105B: 105B parameters; ~9B activated per token; 128,000-token context window
- Government support: ₹246.72 crore under IndiaAI Mission's Innovation Centre
- IndiaAI Mission budget: ₹10,300 crore; GPU pool: 38,000+ (expanding to 58,000+)
- Languages: 22 scheduled Indian languages + English
- Summit recognition: Amit Shah — "exemplifies why the future belongs to India"; PM Modi publicly lauded
- BharatGen Param2 (comparator model): 17B parameters; ₹988.6 crore funding; 22 Indian languages
- BHASHINI: National language AI platform that Sarvam models feed into