Gnani Unveils Sovereign AI Stack Artha With Open-Weight Model and Enterprise Agents

Gnani, a Bengaluru startup, unveiled its sovereign AI stack Artha, including the 30-billion-parameter open-weight model Evon v3.3 and the Plexus agent-building platform, at the official residence of Vice President C P Radhakrishnan in New Delhi. The launch, part of the India AI Mission with an initial outlay of Rs 10,372 crore, aims to provide Indian enterprises and public institutions with domestic AI infrastructure. CEO Ganesh Gopalan said Evon v3.3 outperformed unnamed models on the MILU benchmark and uses about 20 percent fewer tokens per Indian-language word than GPT-5's tokenizer.

August 28, 2026 , Inside AI — A Bengaluru startup has released a foundational language model trained from scratch and an agentic AI platform, both aimed at Indian enterprises and public institutions. Gnani unveiled Evon v3.3 , a 30-billion-parameter model, and Plexus , an agent-building platform, under its new sovereign AI stack called Artha . The launch took place at the official residence of Vice President C P Radhakrishnan in New Delhi. The move signals a shift in India's AI strategy from consuming foreign models to building domestic infrastructure. Gnani is one of 12 ventures selected under the India AI Mission , a government program with an initial outlay of Rs 10,372 crore . Radhakrishnan framed sovereign LLMs as India's greatest advantage. He stressed that India's approach should focus on making AI open, affordable, and accessible so innovation uplifts society as a whole. "Both these platforms reflect the growing strength of India's technology ecosystem. This initiative shows that our engineers have the capability not only to use frontier technologies, but also to build them," the vice president said. Evon v3.3 uses a mixture-of-experts architecture, activating roughly 3.5 billion parameters per token. It supports over 11 Indian languages and its weights are available under the Apache 2.0 license on Hugging Face https://huggingface.co . Gnani claims a key efficiency advantage: the model needs about 20 percent fewer tokens per Indian-language word than the tokenizer used by the GPT-5 family. It requires less than half the tokens of byte-level tokenizers in DeepSeek , Llama , and Qwen . The company rebuilt the tokenizer to better support Indian scripts. Token consumption has drawn scrutiny as enterprises face rising AI costs. This has boosted the popularity of Chinese open-weight models like Moonshot AI's Kimi K3 , which reportedly matches leading US models at lower cost. Ganesh Gopalan , CEO and co-founder of Gnani, said Evon v3.3 outperformed a similarly sized unnamed model and an unnamed 105-billion-parameter Indic model on the MILU benchmark. That benchmark spans eleven languages and dozens of academic and professional subjects. Gopalan defined sovereign AI as having control of one's own data and AI models. He said, "Sovereign AI is not about keeping the world out. It is about India having the capability to build for itself - and then for every country that shares its problems." Gnani is targeting banks, insurers, and government bodies that need to meet data residency requirements. Evon v3.3 is designed for self-hosting on a single node, letting organizations keep sensitive customer data within their own infrastructure. Plexus Turns Prompts Into Autonomous Agents Plexus lets enterprise customers build and deploy AI agents through natural language prompts. The agents support tool calling and can work autonomously across documents, systems, and conversations. Customers can choose from a set of underlying models, including Evon v3.3. In a pre-recorded demo, Gnani showed how Plexus could design an AI agent that fetches a customer's PAN card details with a single prompt. Another demo built and deployed a swarm of agents for welfare beneficiary programs and grievance resolution by government agencies. Gnani described each agent as a discrete, identity-bearing unit, closer to an employee than a script. Agents combine into workflows built around a defined outcome. An orchestration layer can be human-in-the-loop or AI-led, with guardrails, observability, and audit logging built in rather than added afterwards. India's Sovereign AI Push Gains Momentum The India AI Mission was approved by the Union Cabinet in 2024 to build sovereign AI capabilities, including foundational LLMs, large-scale compute infrastructure, and AI applications for public use. In early 2025 , the government operationalized GPU subsidies and invited startups to apply for compute support. At the India AI Impact Summit in February, Gnani launched Vachana TTS , a text-to-speech model that clones human voices across 12 Indian languages using under 10 seconds of reference audio. The Artha stack positions Gnani among a small group of Indian firms building full-stack sovereign AI. Competitors include Tech Mahindra and an IIT-B consortium, both selected to build domestic AI models under the same mission. Industry analysts note that open-weight models like Evon v3.3 could reduce dependence on foreign APIs. But adoption will depend on real-world performance, support for enterprise workflows, and the cost of self-hosting at scale.