Bengaluru-based Sarvam AI launched its largest language model yet, Sarvam-105B, at the
India-AI Impact Summit 2026, marking a major step in India’s push to build sovereign artificial intelligence systems. The new model, with 105 billion parameters, is designed for complex reasoning, coding, and enterprise use cases, and was trained from scratch using domestic compute infrastructure under the government’s IndiaAI Mission.
The launch is significant because it addresses a key gap in India’s AI ambitions - that of building foundational models independently rather than adapting foreign ones. Sarvam said the model supports all 22 official Indian languages, handles long documents with a 128,000-token context window, and uses mixture-of-experts architecture to improve efficiency.
The company positioned the model as suitable for government deployments, enterprise automation, and developer applications, particularly in multilingual and voice-first environments.
What makes Sarvam-105B technically different from its earlier models?
The biggest difference lies in how Sarvam’s new model was built.
Sarvam-105B was trained from scratch, unlike Sarvam-M, the company’s earlier model which was released in May 2025. It was built on top of Mistral Small, a base model developed by French AI firm Mistral. Sarvam-M had about 24 billion parameters and was fine-tuned to improve Indian language capabilities and reasoning performance.
However, the new 105B model represents a shift toward foundational independence. It uses mixture-of-experts architecture, which activates only a fraction of parameters during each task, improving computational efficiency without reducing performance.
It also supports much longer inputs with a 128,000-token context window, through which the model can analyse large documents such as balance sheets, legal filings, or technical manuals in a single prompt. Sarvam said this allows enterprises and government agencies to automate complex workflows more efficiently.
The model was trained using government-supported infrastructure under the IndiaAI Mission, with hardware support from Nvidia and data centre capacity provided by Yotta.
Check India AI Summit 2026 Day 4 Updates Why did Sarvam’s earlier model face criticism?
When it was released in May 2025, Sarvam-M was positioned as a major milestone for India’s AI ecosystem. But it quickly drew criticism because it represented a foreign foundation model and the developers fine-tuned it with Indian datasets. While this improved its performance on Indian languages and reasoning benchmarks, the reliance on overseas architecture raised questions about whether it qualified as a sovereign AI model.
The criticism reflected broader concerns as India’s AI strategy emphasises building domestic foundation models to reduce dependence on foreign providers, particularly for government and strategic applications.
Developers and industry observers argued that adapting foreign models did not address the issue of structural dependence on external AI infrastructure.
Sarvam-105B directly responds to that criticism by moving to a fully domestically trained model.
How does Sarvam-105B compare with global models like ChatGPT and DeepSeek?
Sarvam has positioned the new model as competitive within its size category, particularly among open-source and mid-scale frontier models.
The company said Sarvam-105B performs strongly in reasoning, coding, and multilingual benchmarks, and outperformed several open models in its class. It also
claimed the model surpassed China’s DeepSeek-R1, a much larger 600-billion-parameter model, on certain benchmarks while using fewer active parameters due to its mixture-of-experts design.
Sarvam executives said the model performs on par with several open and closed models of similar scale.
However, the largest global models remain significantly bigger as systems such as Meta’s Llama 3.1 reach up to 405 billion parameters, while proprietary models such as OpenAI’s GPT-series and Google’s Gemini operate at frontier scale with massive training infrastructure.
Sarvam’s competitive advantage lies less in absolute size and more in localisation, efficiency, and multilingual optimisation.
The model is designed specifically to handle 22 Indian languages, including code-mixed formats such as 'Hinglish', which global models often struggle to interpret accurately.
Sarvam has raised over $50 million till now and is backed by Lightspeed, Peak VX Partners, and Khosla Ventures. The AI firm is valued at roughly $200 million which is a fraction of OpenAI’s $500 billion valuation, yet in a market shaped by linguistic diversity and the push for sovereign AI infrastructure, that difference carries less weight than it would in more uniform ecosystems.
Why this launch matters for India’s sovereign AI ambitions
The previous AI system depended on foreign model alterations but the new system demonstrates that local businesses can develop extensive language models by using their existing computing resources and their own data collections.
This indicates an upcoming reduction in reliance on foreign AI providers and gives governments and enterprises greater control over deployment, security, and data governance.
The model’s release also aligns with the IndiaAI Mission’s goal of building domestic AI infrastructure across compute, models, and applications.
Sarvam-105B does not yet match the scale of the world’s largest models but it marks one of the first attempts by an Indian start-up to build a frontier-scale language model independently.
 "Why Sarvam's new 105B model marks a shift in India's sovereign AI ambitions")
