Loading…
Vice President C.P. Radhakrishnan launched 'Gnani Artha', an indigenous sovereign artificial intelligence stack, in New Delhi on 28 August 2026.
The stack has two parts: Evon 3.3, a 30-billion-parameter language model, and Plexus, an agentic platform for building and running AI agents.
Evon 3.3 uses a mixture-of-experts design that activates only about 3.5 billion parameters for any given token, and was trained on about 2 trillion tokens.
It supports more than 11 Indian languages and its weights are released under the Apache 2.0 open licence, with 70-billion and 100-billion parameter models planned.
The stack is built by Gnani.ai, a Bengaluru voice-AI company founded in 2017 and selected under the IndiaAI Mission in 2025 to build a multilingual real-time voice model.
Two technical ideas carry this story. The first is sovereign AI: a stack whose model weights, training and deployment sit within the country, so that an enterprise or a government department can run it on its own hardware and keep the data inside Indian jurisdiction instead of sending it to a foreign cloud API. Releasing the weights under Apache 2.0 is what makes that possible in practice, because a licence that permits inspection and modification lets an organisation audit and fine-tune the model rather than merely rent it. The second is the mixture-of-experts architecture. A dense model runs every parameter for every token; a mixture-of-experts model splits its parameters into specialised sub-networks and routes each token to only a few of them. Evon 3.3 holds 30 billion parameters but activates about 3.5 billion per token, so the compute and memory cost of running it is closer to that of a small model while the stored knowledge is that of a large one. That is the specific reason a comparatively small Indian company can put a frontier-style model into production without frontier-scale compute.
Approved in March 2024, it funds compute capacity, datasets and indigenous foundation models; Gnani.ai was among the start-ups selected to build one.
On-premise deployment of an Indian model is a compliance argument as much as a technical one, since it keeps personal data within the operator's own control.
Apache 2.0 weights let anyone inspect, fine-tune and redistribute the model, but a model is fully open source only if training data and code are released too; exam questions often blur this distinction.
Sits alongside BharatGen and Sarvam in the set of indigenous Indic-language model efforts backed under the IndiaAI Mission.