Gnani Launches 30B Evon Model and Self-Hosted Agent Stack for Indian Institutions

Gnani has launched Artha, a sovereign AI stack aimed at Indian enterprises and public institutions. It combines Evon v3.3, a 30-billion-parameter open-weight language model supporting more than 11 Indian languages, with Plexus, a platform for building enterprise AI agents. Evon’s weights are free to download from Hugging Face under the Apache 2.0 license.

The distinction is the package. Open weights give developers access to Evon itself, while Plexus is intended to turn a chosen underlying model, including Evon v3.3, into self-hosted agents that can work across documents, systems, and conversations. That places model availability and enterprise deployment in the same product offer rather than treating them as separate purchases or engineering projects.

The stated edge is in how Indian-language text is counted

Gnani’s performance argument centers on Evon’s rebuilt tokenizer, the component that breaks text into the units a model processes. The company says it uses about 20% fewer tokens per Indian-language word than the GPT-5 family tokenizer, and less than half as many as byte-level tokenizers used by DeepSeek, Llama, and Qwen. If those comparisons hold for a customer’s workload, fewer tokens could reduce the volume of model processing required for that text.

The deployment contrast is control over data

Artha’s other claim is operational rather than linguistic. Evon v3.3 is designed to run on a single node in a self-hosted deployment, which Gnani says lets organizations retain sensitive customer data within their own infrastructure. Plexus extends that approach to agents: users can create and deploy them from natural-language prompts, with tool calling for actions across company materials and systems.

What the stack puts together

• Evon v3.3 supplies the 30-billion-parameter, open-weight foundation model and support for more than 11 Indian languages.
• Plexus supplies the layer for building and deploying self-hosted agents with natural-language prompts and tool calling.
• The single-node design is intended to support deployments in which sensitive customer data remains inside an organization’s infrastructure.

A local-model effort backed by a national program

Gnani is one of 12 local entities selected under the India AI Mission to develop sovereign AI capabilities. The program was approved in 2024 with an outlay of Rs 10,372 crore, according to the launch account. Artha therefore arrives as both a commercial product and an example of the domestic capability-building effort associated with that mission.

The launch sets up a practical choice for organizations with Indian-language workloads: use an open model whose weights can be downloaded, or use the broader stack to keep model serving and agent operations under local control. Gnani’s token-efficiency figures are company claims, while the release’s concrete proposition is the combination of those claims with open licensing, single-node self-hosting, and an agent layer built for enterprise systems.