Current Affairs Topics Archive
International Relations Economics Polity & Governance Environment & Ecology Science & Technology Internal Security Geography Social Issues Art & Culture Modern History

Bengaluru firm unveils two AI language models

February 18, 2026 4 min read Science & Technology GS Paper 3
On This Page ▼

What Happened

  • Bengaluru-based AI startup Sarvam AI launched two large language models -- Sarvam-30B (30 billion parameters) and Sarvam-105B (105 billion parameters) -- at the India AI Impact Summit 2026 in New Delhi on 18 February 2026.
  • Sarvam-30B was pre-trained on 16 trillion tokens with a 32,000-token context length; Sarvam-105B supports a 128,000-token context window for enterprise-grade workloads.
  • CEO Pratyush Kumar stated that both models were built entirely from scratch with no external data dependency.
  • The models were trained using computing resources provided under the government-backed IndiaAI Mission, with infrastructure support from data centre operator Yotta and technical support from Nvidia.
  • Sarvam-30B's benchmarks in general reasoning and coding are on par with models such as Gemma 27B, Mistral-32-24B, Qwen-30B, and Nemotron-30B; Sarvam-105B competes with OpenAI's GPT-OSS-120B and Alibaba's Qwen-3-Next-80B.
  • Both models will be released as open source, targeting developers, enterprises, and government agencies.
  • The company also demonstrated a multilingual AI chatbot named "Vikram" (named after physicist Vikram Sarabhai) that runs on basic feature phones and converses in Hindi, Punjabi, and other Indian languages.

Static Topic Bridges

Sovereign AI and India's National AI Mission

Sovereign AI refers to a country's ability to develop and control its own artificial intelligence infrastructure -- including models, training data, compute resources, and governance frameworks -- without dependence on foreign technology ecosystems.

  • The IndiaAI Mission, launched by the Government of India, aims to build indigenous AI infrastructure, support research, and promote adoption across sectors such as agriculture, healthcare, and education.
  • IndiaAI Mission 2.0 was unveiled at the India AI Impact Summit 2026, marking a shift from infrastructure-building to deeper R&D, widespread adoption, and extending sovereignty to chips, infrastructure, and control systems.
  • The government projects over USD 200 billion in AI-related investments over two years, covering hardware, models, data centres, and applications.
  • India's AI Governance Guidelines, anchored by "seven sutras," provide a principle-based framework for safe and trusted AI innovation aligned with the Viksit Bharat 2047 vision.
  • India is also part of global discussions on AI governance at forums such as the G20 and the Paris AI Summit.

Connection to this news: Sarvam AI's models were trained using IndiaAI Mission compute resources, making them a direct product of India's sovereign AI push. The decision to open-source the models aligns with the government's "AI for All" strategy to democratise access.

Large Language Models (LLMs) -- Technical Foundations

Large Language Models are deep neural networks trained on massive text corpora to understand and generate human language. They form the foundation of modern AI applications including chatbots, translation, summarisation, and code generation.

  • LLMs are typically based on the Transformer architecture (introduced in 2017), which uses self-attention mechanisms to process sequential data.
  • Model size is measured in parameters (learnable weights); larger models generally have greater capability but require more compute. Sarvam's 30B and 105B models represent medium and large-scale LLMs respectively.
  • Training involves pre-training on trillions of tokens (words/subwords) from text data, followed by fine-tuning for specific tasks.
  • Context length (how much text the model can process at once) is a key differentiator: 32K tokens for Sarvam-30B and 128K tokens for Sarvam-105B.
  • Open-source LLMs (such as Meta's Llama, Mistral, and now Sarvam) allow developers to inspect, modify, and deploy models without vendor lock-in, unlike proprietary models from companies like OpenAI or Google.

Connection to this news: Sarvam AI's models are significant as the first Indian-built LLMs trained from scratch (not fine-tuned on existing open-source models) and optimised for Indian languages. Their benchmark performance matching established global models demonstrates India's growing capability in foundational AI research.

Digital India and Emerging Technology Ecosystem

India's technology policy framework has evolved from digitisation of government services to building indigenous capability in frontier technologies such as AI, quantum computing, and semiconductors.

  • The Digital India programme (launched 2015) laid the groundwork for digital infrastructure, including Aadhaar, UPI, and DigiLocker.
  • IndiaAI Mission 2.0 proposes a UPI-style AI platform offering ready-to-use, vetted AI solutions to help MSMEs improve productivity.
  • India's startup ecosystem is the third largest globally, with over 100 unicorns, and AI startups are a growing segment.
  • Sarvam AI was founded in 2023 by Vivek Raghavan (digital public infrastructure background) and Pratyush Kumar (open-source AI for Indian languages). It has raised USD 53.8 million from Lightspeed Venture Partners, Peak XV Partners, and Khosla Ventures.
  • The company is a member of the AI Alliance, a global consortium co-led by Meta and IBM promoting open-source AI.

Connection to this news: Sarvam AI's demonstration of its chatbot "Vikram" running on a basic feature phone illustrates the potential for AI democratisation in India, where hundreds of millions still use feature phones. This aligns with the Digital India vision of inclusive technology access.

Key Facts & Data

  • Sarvam-30B: 30 billion parameters, trained on 16 trillion tokens, 32K token context length
  • Sarvam-105B: 105 billion parameters, 128K token context window
  • Both models trained from scratch using IndiaAI Mission compute resources (Yotta data centre, Nvidia support)
  • Both to be released as open source
  • Sarvam AI founded: 2023, by Vivek Raghavan and Pratyush Kumar
  • Total funding raised: USD 53.8 million (Lightspeed, Peak XV, Khosla Ventures)
  • IndiaAI Mission 2.0: projects over USD 200 billion in AI investments over two years
  • Benchmarks: Sarvam-30B on par with Gemma 27B, Mistral-32-24B, Qwen-30B; Sarvam-105B competes with GPT-OSS-120B, Qwen-3-Next-80B
  • "Vikram" chatbot demonstrated on basic feature phones in Hindi and Punjabi