Gnani.ai launches Artha, a sovereign AI stack for India

A new AI stack appeared in New Delhi with a distinctly local starting point. Gnani.ai unveiled Artha, combining its 30 Bn-parameter Evon v3.3 model with Plexus, a platform for building workflow-specific AI agents. The launch was presented by India’s vice president, CP Radhakrishnan, as the Bengaluru startup targets enterprises and public institutions.

Evon was trained natively across 11 Indian languages and on 2 Tn tokens, with an emphasis on reasoning and Indic-language performance. Its weights — the learned patterns that shape how a model interprets inputs and generates responses — are available upon request on Hugging Face under an Apache 2.0 licence. Plexus adds the operational layer: organisations can use the model to build and deploy agents for tasks such as underwriting, advertising and payments reconciliation while retaining control of their data and technology infrastructure.

Gnani.ai’s efficiency claim is central to the pitch. The company says Evon consumes around 40% fewer tokens than comparable models for Indian-language workloads, potentially lowering compute costs when AI moves beyond demonstrations. The company says its focus is production use rather than benchmark performance, but the evidence remains early.

The first customer signals are tangible but not deployments. Gnani.ai says five of 20 enterprises at a recent meeting in Pune expressed interest and began building on Evon. The engagements are still at an early stage, and no Evon deployment has gone live. In underwriting and payments reconciliation, humans will initially remain involved; more routine work, such as matching documents and data during loan onboarding, is expected to become increasingly autonomous.

So what changes in practice? Gnani.ai says the stack is designed to address data sovereignty, deployment costs at scale, and the ability to work across Indian languages and real-world use cases. The next steps are still plans: Gnani.ai is considering 70 Bn- and 100 Bn-parameter models, wants to expand from 11 to 22 languages, and has not disclosed a timeline for its speech-to-speech technology. Artha is therefore best understood as an early customer platform, not a proven large-scale deployment.