Sarvam AI’s Indigenous LLM Revolution: The Future of Vernacular AI in India

Explore how Sarvam AI is building India’s first sovereign LLMs to power vernacular AI and digital inclusion across 22 Indian languages.

Sarvam AI, a Bengaluru-based startup founded in July 2023 by Vivek Raghavan and Pratyush Kumar, is spearheading India’s ambition to develop sovereign artificial intelligence (AI) through indigenous foundation models. With a mission to create AI that resonates with India’s linguistic and cultural diversity, Sarvam AI is building large language models (LLMs) from scratch, optimized for 10 major Indian languages alongside English. This article delves into Sarvam AI’s transformer-based architecture, training methodologies, advancements in vernacular natural language processing (NLP), real-world use cases, and the broader implications for India’s AI startup ecosystem. As of May 3, 2025, Sarvam AI’s efforts, including its flagship models Sarvam-1 and Sarvam 2B, position it as a pivotal player in India’s AI landscape.

Sarvam AI’s Mission and Strategic Context

India’s AI ecosystem has historically leaned on adapting Western models, but Sarvam AI is charting a new course by developing foundation models tailored to India’s needs. The startup’s vision aligns with the IndiaAI Mission, a ₹10,372 crore government initiative launched in 2024 to bolster domestic AI innovation. Sarvam AI was selected in April 2025 to build India’s first sovereign LLM, leveraging 4,096 Nvidia H100 GPUs for six months to create a 70-billion-parameter model capable of advanced reasoning and voice-first interactions across 22 Indian languages. This move underscores India’s push for strategic autonomy in AI, addressing challenges like linguistic diversity, data scarcity, and global competition from players like OpenAI and China’s DeepSeek.

Sarvam AI’s focus on vernacular intelligence is critical in a country where digital adoption is surging but literacy barriers persist. With 22 official languages and hundreds of dialects, India’s linguistic diversity demands AI models that can process and generate text and speech in languages like Hindi, Odia, Tamil and Telugu,. Sarvam’s models aim to bridge this gap, enabling applications in customer service, education, healthcare, and governance. The startup’s full-stack approach—encompassing data curation, model training, and deployment—sets it apart from peers that fine-tune existing models.

Transformer Architecture: The Backbone of Sarvam’s Models

Sarvam AI’s foundation models, including Sarvam-1 and Sarvam 2B, are built on transformer architectures, the industry standard for LLMs since Google’s seminal 2017 paper, “Attention is All You Need.” Transformers rely on self-attention mechanisms to weigh the importance of different words in a sequence, enabling context-aware language understanding and generation. Sarvam’s models, with 2 billion parameters each, are classified as small language models (SLMs) compared to behemoths like GPT-4, which has over a trillion parameters. However, their efficiency and specialization for Indian languages make them highly effective.

Architectural Optimizations

Sarvam-1 and Sarvam 2B employ a custom tokenizer optimized for Indic scripts, reducing tokenization costs compared to English-centric models. Tokenization, the process of breaking text into smaller units (tokens), is computationally expensive for Indic languages due to their complex scripts and limited training data. Sarvam’s tokenizer achieves rates of 1.4 to 2.1 tokens per word. This optimization enhances inference speeds, making Sarvam-1 4–6 times faster than  models like Llama-3.1-8B and Google’s Gemma-2-9B on Indic language tasks.

The models use a transformer architecture with multiple layers of interconnected nodes, incorporating multi-head self-attention and feed-forward neural networks. Sarvam likely employs techniques like layer normalization and residual connections to stabilize training and improve convergence. For multilingual tasks, the models are designed to handle cross-lingual transfer, allowing knowledge learned in one language (e.g., English) to enhance performance in others (e.g., Hindi). This is critical for India’s multilingual context, where code-mixing and cross-lingual queries are common.

Model Variants

Sarvam AI is developing three variants under the IndiaAI Mission:

• Sarvam-Large: A 70-billion-parameter model for advanced reasoning and generation, targeting complex tasks like policy analysis and scientific research.

• Sarvam-Small: Optimized for real-time interactive applications, such as chatbots and voice agents, with low latency and high efficiency.

• Sarvam-Edge: A compact model for on-device tasks, enabling AI deployment on smartphones and edge devices with limited compute resources.

These variants leverage shared architectural principles but are fine-tuned for specific use cases, reflecting Sarvam’s multi-scale approach to AI deployment.

Training Techniques: Overcoming Data Scarcity

Training LLMs for Indian languages is challenging due to the scarcity of high-quality, diverse datasets. Sarvam AI addresses this through innovative data curation and synthetic data generation, creating proprietary models like Sarvam-2T (2 trillion tokens) for Sarvam-1 and a 4-trillion-token dataset for Sarvam 2B.

Data Curation with Sarvam-2T

The Sarvam-2T corpus, used for Sarvam-1, comprises 20% Hindi, with the remainder distributed across nine other Indian languages, English, and programming languages. To overcome data scarcity, Sarvam employs synthetic data generation, creating high-quality text using rule-based systems and smaller pre-trained models. This approach ensures diversity and depth, addressing gaps in web-scraped Indic datasets like Sangraha. The corpus is pre-processed to remove duplicates, noise, and low-quality content, using NVIDIA’s NeMo Curator for domain and quality classification.

Sarvam 2B’s 4-trillion-token dataset builds on this approach, with a focus on balancing linguistic representation across 10 Indian languages: Hindi, Tamil, Telugu, Malayalam, Punjabi, Odia, Gujarati, Marathi, Kannada, and Bengali. The dataset’s scale and quality enable Sarvam 2B to outperform larger models on benchmarks like MMLU, ARC-Challenge, and IndicGenBench, achieving 86.11% accuracy on TriviaQA compared to Llama-3.1-8B’s 61.47%.

Training Infrastructure

Sarvam leverages NVIDIA’s NeMo framework and H100 Tensor Core GPUs on Yotta’s Shakti Cloud for training. Sarvam-1’s training, completed in five days using 1,024 GPUs, demonstrates the efficiency of this setup. The NeMo framework optimizes data pipelines and model training, incorporating techniques like mixed-precision training and gradient checkpointing to reduce memory usage and accelerate convergence. For inference, Sarvam uses NVIDIA’s TensorRT-LLM with FP8 precision, enhancing deployment efficiency on H100 GPUs and edge devices.

Fine-Tuning and Specialization

Post-training, Sarvam fine-tunes its models for specific tasks, such as translation, speech recognition, and document parsing. Techniques like supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF) are likely employed to align models with user expectations and improve performance on vernacular tasks. For example, Sarvam’s Shuka v1, an audio understanding model, is fine-tuned for spoken Hindi, enabling applications in voice-based customer service.

Vernacular NLP: Strengths and Applications

Sarvam AI’s models excel in vernacular NLP, addressing India’s unique linguistic challenges. Key strengths include:

Multilingual Proficiency

Sarvam-1 and Sarvam 2B support 10 Indian languages plus English, handling monolingual and multilingual tasks like translation, summarization, and question answering. Their cross-lingual capabilities, evaluated on IndicGenBench, enable seamless transitions between languages, crucial for India’s code-mixing culture.

Voice-Enabled AI

Sarvam’s platform emphasizes voice-first interfaces, critical for a country with low literacy rates. Sarvam Agents, voice-enabled multilingual bots, are deployed via phone, WhatsApp, or in-app interfaces, starting at ₹1 per minute. Applications include customer service for companies like Zepto and Zomato, government-citizen interactions, and financial services. Shuka 1.0, an open-source AudioLM, supports accurate voice-to-text conversion in Indian languages.

Efficiency and Scalability

The models’ compact size and optimized tokenization make them suitable for edge devices, reducing dependency on cloud infrastructure. This scalability supports applications in rural areas with limited connectivity, enhancing digital inclusion.

Real-Life Use Cases in Progress

As of May 3, 2025, Sarvam AI’s models are powering transformative applications across India. In e-commerce, Zepto and Zomato deploy Sarvam Agents for multilingual customer support, handling Hindi and Tamil queries via WhatsApp and phone, reducing response times by 40%. In healthcare, Apollo Hospitals is piloting Sarvam’s voice-enabled AI to assist rural patients in scheduling appointments and accessing medical advice in regional languages like Telugu and Kannada, improving access for non-English speakers. The Government of Odisha uses Sarvam-1 for a citizen engagement platform, enabling Odia-speaking residents to interact with public services through voice and text. In education, edtech startup Byju’s integrates Sarvam 2B to provide personalized tutoring in Marathi and Gujarati, enhancing learning outcomes for vernacular-medium students. These use cases demonstrate Sarvam’s ability to address India’s diverse needs, from urban commerce to rural governance.

Implications for India’s AI Startup Ecosystem

Sarvam AI’s advancements have far-reaching implications for India’s AI startup ecosystem, which raised $166 million in 2024, down from $518.2 million in 2022. Key impacts include:

Catalyzing Innovation

Sarvam’s open-source models, available on Hugging Face, empower developers to build Indic language applications, from chatbots to legal workbenches like A1. This fosters a vibrant ecosystem of AI-driven startups, particularly in sectors like legaltech, fintech, and edtech.

Economic Transformation

By automating customer-facing roles in industries like banking, telecom, and quick commerce, Sarvam’s models reduce costs and improve efficiency. For instance, replacing human agents with AI bots could save millions for large enterprises, driving economic productivity.

Global Competitiveness

Sarvam’s sovereign AI approach positions India to compete with global leaders like OpenAI and DeepSeek. By leveraging local talent and infrastructure, Sarvam reduces reliance on foreign models, enhancing data security and national sovereignty.

Challenges Ahead

Despite its progress, Sarvam faces hurdles, including talent retention in a competitive global market, interoperability with diverse platforms, and scaling high-performance computing infrastructure. India’s contribution to global AI research remains under 1.5%, highlighting the need for academic and industry collaboration.

Sarvam AI is redefining India’s AI landscape by building indigenous foundation models tailored to its linguistic and cultural diversity. Its transformer-based architecture, innovative training techniques, vernacular NLP strengths, and real-world applications enable transformative impacts across sectors. As India’s first sovereign LLM takes shape under the IndiaAI Mission, Sarvam AI is poised to drive economic growth, foster innovation, and position India as a global AI contender. For businesses, developers, and policymakers, Sarvam’s journey offers a blueprint for leveraging AI to address India’s unique challenges and opportunities.