Gnani’s 30B Model Is Taking Aim At Sarvam’s 105B

For years, Indian enterprises have faced a choice: use powerful foreign AI models or build more of their AI stack locally.

The IndiaAI Mission is trying to change that. Twelve organisations and consortia have now been selected to develop indigenous foundation models and large or small language models. They include Sarvam AI, Soket AI, Gnani AI and the IIT Bombay-led BharatGen consortium.

Gnani is taking a different route from the companies building very large general-purpose models.

The Bengaluru-based voice AI company recently launched Gnani Artha, a sovereign AI stack that combines its Evon v3.3 language model with Plexus, an agentic AI platform for enterprise and government workflows. Vice President C.P. Radhakrishnan unveiled the stack in New Delhi on August 28.

At the centre of Artha is Evon v3.3, a 30-billion-parameter mixture-of-experts model. Around 3.5 billion parameters are active for a given token. The model supports 11 languages, including Hindi, Bengali, Telugu, Tamil, Marathi, Gujarati, Kannada, Malayalam, Odia and Punjabi, along with English. It has a 128K context window and is available on Hugging Face under an Apache 2.0 licence.

Gnani is positioning Evon less as another frontier model and more as infrastructure for enterprises that want to build AI around their own data.

That distinction is central to the company's strategy.

“We are an enterprise-focused company, and we work with 200 organisations in about five countries,” Gnani CEO and co-founder Ganesh Gopalan tells AIM. “Our focus is the reason for launching this also is fairly simple… taking an open source, open weight model rather than a broader closed frontier model.”

The argument is straightforward. Enterprises may not always need the biggest model available. They may need a model that performs well on their languages, runs at a manageable cost and can be deployed within infrastructure they control.

The 105B Model Challenge

Gnani is making an aggressive claim with Evon v3.3.

The company says the model outperforms Sarvam AI’s 105-billion-parameter model on MILU, an Indic-language understanding benchmark, across 10 of the 11 languages tested. It also says Evon outperforms Sarvam's 30B model across all 11 languages.

These are company-reported results, rather than independently verified findings. Gnani says its testing covered dozens of benchmarks and that Evon is available for external evaluation.

The comparison is still significant. It challenges the assumption that a larger parameter count automatically translates into better performance on Indian languages.

But model size alone is not the point Gnani is trying to make.

Evon is a mixture-of-experts model. While it has 30 billion parameters in total, only about 3.5 billion are activated for each token. This reduces the computational cost per inference compared with a dense model of similar size.

For enterprises, Gnani argues, the more relevant question is not simply how large a model is. It is how well it performs on a specific workload and how much it costs to run.

“We are not doing research projects. We want to do a model that enterprises will pick and make ROI for them,” Gopalan says.

Fewer Tokens, Lower Costs

Tokenisation is another part of Gnani's pitch.

The company claims Evon uses about 20% fewer tokens per Indian-language word than the tokeniser used by the GPT-5 family. It also claims to use less than half the tokens required by byte-level tokenisers used by models such as DeepSeek, Llama and Qwen.

Gnani says the gains vary by language. It claims Evon uses 37% to 42% fewer tokens for languages such as Gujarati and Malayalam compared with certain competing models.

These figures are also based on the company's own testing and should be treated as such.

The underlying idea, however, is important for enterprise AI. Fewer tokens can reduce inference costs and allow applications to process more information within a fixed context window.

That matters when an AI system is handling millions of interactions.

“You’re doing it because you want to make something work, not a POC, but you want to make something work in production,” Gopalan says. “Fine, you'll get great accuracy [in research], but will it make money for you?”

Sovereignty Means Control

Evon is available as an open-weight model. Gnani says it can be deployed within an organisation's own data centre or virtual private cloud.

That gives enterprises more control over where their data is processed. It also allows them to customise the model for specific workloads.

This is where Artha's second component, Plexus, comes in.

Plexus is an agentic AI platform that allows enterprises to build workflows involving multiple AI agents. Potential applications include government grievance handling and banking reconciliation.

For example, one agent could process a citizen complaint. Another could classify it. A third could route it to the relevant department.

But Gnani is not pitching full autonomy for every use case.

Gopalan says regulated industries will continue to require human oversight.

“Having a simplistic notion that all agents will be 100% autonomous, you are going to have orchestrator agents with about a million agents that work for that goal may not be true for every industry,” he says.

That caution is important as enterprises begin experimenting with agentic systems that can access internal data and take actions on employees' behalf.

Compute Is Still The Constraint

Gnani's launch also highlights a problem facing India's sovereign AI ambitions: compute.

Training and running large models requires access to GPUs. For smaller companies, that can be a significant constraint.

“The key constraint has been GPUs and it will be GPUs,” Gopalan says. “Especially today when it is really so tough to get hold of any form of compute.”

He credits government-backed access to compute with helping Gnani develop its models.

The IndiaAI Mission has expanded access to GPU infrastructure for startups, researchers and other users. The government said in 2025 that more than 34,000 GPUs had been provisioned through the IndiaAI Compute Portal.

But Gopalan expects the next challenge to be inference efficiency.

“The future is going to be a bunch of different things. One is the quality of your model. But what is going to be equally important is the ability of your model to run on various inferencing chips,” he says.

Gnani is currently targeting 11 Indian languages with Evon and wants to eventually expand to all 22 scheduled Indian languages.

The company also sees a market beyond India. Gopalan says it has received interest from countries in Africa, the Middle East and East Asia that face similar questions around local languages, data control and dependence on foreign AI systems.

From Text To Voice

Evon is not Gnani's final destination.

The company was selected under the IndiaAI Mission to build a 14-billion-parameter voice AI foundation model for multilingual, real-time speech processing. The government's programme describes it as a model with advanced reasoning capabilities.

Gnani has previously said the voice model is intended to support real-time speech-to-speech applications and eventually expand across India's 22 scheduled languages.

For Gnani, Evon provides the language foundation for that larger ambition.

The company is therefore making two bets at once.

The first is that Indian-language AI does not always need a massive model to compete on specific workloads.

The second is that enterprises will increasingly care about where their AI runs, how much it costs and how much control they have over it.

Evon v3.3 is an early test of that thesis. The benchmark claims will need to survive independent scrutiny. But if the model can deliver comparable performance at a fraction of the compute cost, Gnani may have found a more practical way to sell sovereign AI to enterprises.