India’s home-grown AI race has a new heavyweight. Bengaluru-based Gnani AI has launched Evon 3.3, a 30-billion-parameter open-weights language model built for Indian languages, alongside its agentic AI platform Plexus under a sovereign AI stack called Artha. Vice President C. P. Radhakrishnan unveiled the stack in New Delhi on August 28, 2026.
The launch lands at a time when India is pushing harder to build its own foundational models, cheaper compute and local-language AI infrastructure. Evon 3.3 will not remove foreign technology from India’s AI chain overnight. Still, its self-hosting, Indic-language focus and open weights could give banks, insurers, government departments and developers another route beyond depending entirely on overseas LLM APIs.
What Makes Gnani Evon 3.3 Different From A Typical Global LLM?
Evon 3.3 uses a mixture-of-experts design with 30 billion total parameters, but only about 3.5 billion are active for each token. Gnani says that helps reduce the computing needed during inference. The model supports English plus Hindi, Bengali, Telugu, Tamil, Marathi, Gujarati, Kannada, Malayalam, Odia and Punjabi, with a context window of up to 128K tokens.
The model is available under the Apache 2.0 licence, with weights offered through Gnani’s official Hugging Face repository. Gnani has also rebuilt the tokenizer for Indian scripts. According to the company, Evon 3.3 uses roughly 20% fewer tokens per Indian-language word than the tokenizer used by the GPT-5 family.
That detail could prove more commercially important than the parameter count. Indian scripts can become expensive when tokenized inefficiently. Fewer tokens can mean lower inference bills, faster responses and more text fitting into the same context window.
Key points from the launch include:
- 30 billion total parameters, with roughly 3.5 billion active per token.
- Native training focus across 11 languages, including ten Indian languages plus English.
- Open weights under Apache 2.0, with self-hosting supported on enterprise infrastructure.
- More than 2 trillion training tokens were used across Indic and domain-specific data, according to Gnani executives.
- Plexus adds agentic workflows, connecting models with tools, documents and internal systems.
Why This Launch Arrives At A Big Moment For India’s AI Push
Evon 3.3 is not appearing in isolation. Gnani is one of 12 organisations and consortia selected under the IndiaAI Mission’s indigenous foundational-model programme. The government says more than 38,000 high-end GPUs have been onboarded for shared compute, while AIKosh has grown into a national platform for datasets and AI models.
The Indian AI field is also getting busier. Sarvam, another sovereign-AI company backed by the IndiaAI effort, announced a $300 million Series B in June 2026. This week, Vahan AI also disclosed that it had fine-tuned Nvidia’s 30-billion-parameter Nemotron 3 Nano for voice-led blue-collar recruitment.
That makes Evon’s launch part of a broader shift. Indian companies are no longer only building applications on top of American or Chinese foundation models. More teams are now working on the model, speech, agent and deployment layers themselves.
Gnani’s official launch announcement described Artha as an end-to-end sovereign stack for Indian enterprises and public institutions. Its official Gnani Artha page also puts local deployment and data residency at the centre of the pitch.
Can Evon 3.3 Really Cut India’s Dependence On Foreign LLMs?
Yes, but only in specific layers of the AI stack.
For organisations currently sending prompts, documents or customer data to externally hosted LLM services, an open-weight model that can run inside a private data centre or virtual private cloud changes the equation. Sensitive banking, insurance, government and enterprise workloads can stay under local operational control.
There is also a language advantage to test. Gnani says its MILU results put Evon 3.3 ahead of Sarvam’s 30B model across all 11 tested languages and ahead of Sarvam’s 105B model in 10 of 11. Those are company-reported benchmark results, so independent testing will be crucial before treating them as settled performance rankings.
But “sovereign” does not mean every layer is Indian. Evon 3.3’s published architecture is based on Nvidia’s Nemotron hybrid MoE family, and Gnani told The Economic Times it used about 1,500 Nvidia GPUs across development stages. India can therefore reduce dependence on foreign model APIs while still relying heavily on imported accelerator hardware and global software ecosystems.
The real test is adoption. If Indian enterprises can run Evon cheaply, tune it for their own data and get strong multilingual performance, it becomes a practical alternative. If deployment costs, accuracy or developer tooling lag behind global models, foreign LLMs will remain difficult to displace.
FAQs
What is Gnani Evon 3.3?
It is a 30-billion-parameter open-weight language model focused on English and major Indian languages.
Who launched Gnani Artha?
Vice President C. P. Radhakrishnan unveiled Gnani Artha in New Delhi on August 28, 2026.
How many languages does Evon 3.3 support?
It supports English and ten Indian languages, including Hindi, Tamil, Telugu, Bengali and Marathi.
Is Evon 3.3 completely independent of foreign technology?
No. Its architecture draws on Nvidia Nemotron, while development also relied on Nvidia GPU infrastructure.
Can companies host Evon 3.3 themselves?
Yes. Gnani says enterprises can self-host the open-weight model within private infrastructure for greater control.
India’s AI Race Gets A New Home-Grown Contender
IIT Madras Alum Anima Anandkumar Named In TIME100 AI: How Does Her Weather Tech Work?
