Gnani.ai launches Artha sovereign AI stack

Gnani.ai, a voice AI company, unveiled Artha, a sovereign AI stack, on August 28 in New Delhi. The stack, tailored for Indian companies and public institutions, was presented by Vice President CP Radhakrishnan. Artha enables organizations to run AI models and applications within their own infrastructure, which is crucial for sensitive industries like banking, insurance, and government entities that deal with restricted data and adhere to regulatory guidelines.

Gnani's flagship language model, Evon 3.3, serves as the core of Artha. This model, possessing 30 billion parameters, is fortified by an agentic AI platform called Plexus. Evon 3.3 is released as an open-weights model, while Plexus will be made available to enterprise customers. Developed by pre-training Nvidia's Nemotron model on data specific to Indian languages, Evon 3.3 exhibits enhanced performance in understanding and processing these languages compared to global models.

Specifically, it demonstrates greater efficiency in terms of token consumption and accuracy on language benchmarks such as MILU.

The training corpus for Evon 3.3 comprised over 2 trillion tokens across 11 Indian languages, with an optimized tokenizer that consumes fewer tokens per Indian-language word than tokenizers in models like GPT-5, DeepSeek, Llama, and Qwen. This efficiency translates into reduced costs and latency for enterprises handling large volumes of Indian-language documents or conversations.

The model can run on hardware such as Nvidia's RTX 6000 Pro or L40S, without necessitating high-end accelerators like the B200 or H200. Evon 3.3's codebase is available on Hugging Face under an Apache 2.0 licence, allowing enterprises to build derivative models and deploy them in their own data centers or virtual private clouds.