Gnani AI launches Artha sovereign AI stack with 30-billion-parameter Evon 3.3

Voice AI company Gnani AI on Friday (August 28) launched Artha, a sovereign artificial intelligence stack built around its 30-billion-parameter Evon 3.3 language model and an agentic AI platform called Plexus.

The stack, aimed at Indian companies and public institutions, was unveiled in New Delhi by Vice President CP Radhakrishnan.

According to the company, Artha is designed to allow organisations to run AI models and applications within their own infrastructure, particularly relevant for banks, insurers and government departments that handle sensitive information and face regulatory requirements governing where data is stored and processed.


Gnani added that Evon 3.3 is being released as an open-weights model, while Plexus will be offered to enterprise customers.

Also Read: Business outcomes, not AI models, will decide enterprise deals: Gnani.ai CEO Ganesh Gopalan

Built on Nvidia’s Nemotron

Gnani developed Evon 3.3 by continually pre-training Nvidia’s Nemotron model on its Indic-language and domain-specific data. This was followed by post-training and reinforcement learning.

“We have pre-trained this with our own data, with our own tokens. It is also post-trained with reinforcement learning,” Gnani cofounder and chief executive Ganesh Gopalan told The Economic Times Digital.

“If you compare Evon with the standard Nemotron model, you will see the difference in its Indic-language capabilities, performance, token efficiency and cost. It is significantly more efficient for Indian languages than generic global models in terms of tokens consumed and accuracy on benchmarks such as MILU. You do not see those capabilities in the standard Nemotron model, but you see them in Evon,” he added.

The training corpus contained more than 2 trillion tokens across 11 Indian languages, according to Gnani cofounder and chief product and engineering officer Bharath Shankar.

“We have also optimised Evon for tool calling, language understanding, reasoning, speed and cost,” Shankar said.

Evon 3.3 uses a mixture-of-experts architecture. Although the model has 30 billion parameters, only about 3.5 billion are activated for a given task. Shankar said this allows it to offer reasoning and language-understanding capabilities while consuming less computing power than larger models.

Gnani claims that Evon 3.3 outperforms Sarvam’s 30-billion-parameter model and 105-billion-parameter model in 10 of 11 languages on MILU, a benchmark used to evaluate Indic-language understanding.

“Our model has 30 billion parameters but it outperformed the 105-billion-parameter model in 10 of the 11 languages on which it was trained,” Shankar said.

The results are based on Gnani’s internal testing across about 40-45 benchmarks. The company also tested Evon 3.3 on MMLU and MMLU-Pro, widely used benchmarks for evaluating reasoning and language understanding.

The development process was divided into data cleaning, continual pre-training and post-training, with Gnani using about 1,500 Nvidia GPUs across these stages, Shankar said.

Cutting Indian-language token costs

Gnani has also rebuilt the model’s tokenizer, the component that breaks text into units a model can process, to handle Indian scripts more efficiently.

The company claimed Evon 3.3 requires about 20% fewer tokens per Indian-language word than the tokenizer used by the GPT-5 family and less than half the number used by byte-level tokenizers in models such as DeepSeek, Llama and Qwen.

Also Read: From 100,000 calls a month to 5 million a day: Inside Vobiz’s AI telephony bet

“We evaluate the tokenizer using a measure called token fertility, which looks at how many tokens a model needs to represent text in a particular language. In practice, this means the model consumes fewer tokens to process and generate the same amount of content. It can understand the input language more efficiently and does not need as many tokens to represent it,” Shankar said.

Reducing the number of tokens required to process text can lower inference costs and latency, particularly when organisations are handling large volumes of Indian-language documents or conversations.

Gnani claims that running Evon 3.3 could cost between one-third and one-fifth as much as comparable OpenAI models.

“Our claim of roughly 39% to 40% greater efficiency is based on the tokenizer’s performance. Fewer tokens directly translate into lower costs and reduced latency, making responses faster by a similar margin,” Shankar said.

According to him, Evon 3.3 can run on hardware such as Nvidia’s RTX 6000 Pro or L40S and does not require high-end accelerators such as the B200 or H200, although performance will vary depending on the workload and deployment.

The model weights are available by request on Hugging Face under an Apache 2.0 licence. This will allow enterprises to develop derivative models and deploy them in their own data centres or virtual private clouds.

“We are building purpose-built models in India, initially with a focus on Indian languages, but we will extend them to other languages as well. For us, the sovereign story is about building in India for the world, rather than building only for consumption in India,” Shankar said.

From models to enterprise agents

Plexus forms the second part of the Artha stack. The platform allows enterprises to combine multiple AI agents into workflows, specify when humans must intervene and monitor the agents’ actions.

Also Read: OpenAI-Hugging Face incident exposes cybersecurity’s ‘human-speed’ problem

It can work with different underlying models, including Evon 3.3, and connect AI agents with documents, enterprise software and conversations, the company said.

Potential applications include processing multilingual loan documents, reconciling transactions across bank statements and core banking systems, resolving citizen grievances, underwriting insurance policies and servicing loans, Gopalan said.

Enterprise fine-tuning could become one revenue stream for the open-weights model, alongside Gnani’s agentic platform and applications.

“There are multiple ways in which the commercial angle of launching these open source, open weight models will evolve,” Gopalan said.