Bengaluru-based AI company Gnani AI on Friday launched Gnani Artha, a sovereign AI stack that it says will allow Indian enterprises and government institutions to use AI for tasks involving Indian languages and sensitive data without having to send that data outside their own systems.
Vice-President C P Radhakrishnan unveiled the platform at an event in New Delhi, saying India’s approach to AI should be open, affordable, and accessible.
Gnani Artha brings together two of the company’s existing products — Evon v3.3, a 30-billion-parameter AI model trained across 11 Indian languages, and Plexus, a platform that allows AI agents to carry out tasks across business processes.
The significance of the combination is less about the size of the AI model and more about where and how it can be used. For an Indian bank, for instance, the system could process a loan application containing documents in a regional language, bank statements and other financial records, and use AI to identify inconsistencies. The same capability could be deployed within the bank’s own infrastructure, rather than sending customer data to an external AI service.
This is particularly relevant for banks, insurers and government agencies that deal with large volumes of sensitive information and face requirements around where such data can be stored and processed. Evon v3.3 is available as open weights and can run on a single node, allowing institutions to deploy it in their own data centres or private cloud environments, Gnani said.
The company is also positioning the model as a lower-cost option for processing Indian-language data. It has rebuilt the model’s tokeniser — the system that breaks text into units that an AI model can process — specifically for Indian scripts. Gnani said this allows Evon v3.3 to use about 20% fewer tokens per Indian-language word than the tokeniser used by the GPT-5 family and less than half the tokens required by byte-level tokenisers used by DeepSeek, Llama and Qwen.
That matters because AI models incur computing costs as they process more tokens. Using fewer tokens can therefore reduce the cost of running AI applications, particularly when they are processing thousands or millions of documents or conversations. Gnani said the lower token requirement also enables faster responses and more usable context for each query.
The company said Evon v3.3 has roughly 3.5 billion parameters active at any given time, although the model has 30 billion parameters in total. On MILU, an Indian-language benchmark covering 11 languages, Gnani said the model outperformed a 105-billion-parameter Indic model in 10 of the 11 languages and a similarly sized 30-billion-parameter model across all 11. It also claimed parity with a similarly sized hosted global frontier model.
The other part of Gnani Artha, Plexus, is intended to take AI beyond answering questions and into performing tasks. In government grievance handling, for example, AI agents can capture complaints, identify recurring problems and route cases to the relevant department. In banking, they can match transactions across bank statements, ledgers and payment systems and flag exceptions for human employees to resolve.
Gnani said the workflows can operate with human oversight and include audit logging and other controls. Evon v3.3 weights are available by request on Hugging Face under an Apache 2.0 licence, while Plexus is being made available to enterprise customers.