Gnani Artha isn't trying to become another ChatGPT. Its creators are betting that the biggest mistake in AI is building everything for English first.

For years, the global AI race has looked remarkably simple.

Build a bigger model.

Buy more GPUs.

Train on more data.

Announce a larger parameter count than your competitor.

Then came an uncomfortable question that most of Silicon Valley wasn't particularly interested in solving:

What happens when AI has to work for billions of people who don't primarily think, speak, read, or write in English?

That question sits at the centre of a new AI system launched in India.

On August 28, 2026, Bengaluru-based Gnani AI unveiled Gnani Artha, a sovereign AI stack built around a new language model called Evon v3.3 and an agentic AI platform called Plexus. The launch took place at the Vice President's residence in New Delhi, with Vice President C. P. Radhakrishnan officially unveiling the system.

At first glance, this might sound like another press release announcing another large language model.

It isn't.

Because buried underneath the usual AI buzzwords is something much more interesting.

Gnani's bet is that the future of AI in India might not depend on building the biggest model. It might depend on building one that understands India more efficiently than the models built elsewhere.

And that brings us to one of the most overlooked pieces of AI technology:

The tokenizer.

Yes. The tokenizer.

Something most people have never heard of could determine whether AI becomes affordable enough to serve India at scale.

The AI Race Has a Hidden English Problem

Imagine asking an AI the same question in English and Malayalam.

To a human, they're simply two languages.

To a language model, they can represent very different computational workloads.

Before an AI model can understand text, the text has to be broken into smaller pieces called tokens.

You can think of tokens as the chunks of information a language model processes.

The more tokens required to represent a piece of text, the more computation is generally required to process it.

That means more:

• GPU time
• memory
• latency
• inference cost

For languages that were not heavily prioritised during the development of global AI systems, tokenisation can become surprisingly inefficient.

A sentence may require far more tokens than an equivalent English sentence.

Scale that problem from one conversation to millions of users, and suddenly you have an economic problem.

Not just a language problem.

Gnani says it rebuilt Evon v3.3's tokenizer specifically for Indian scripts. According to the company's published specifications, the model uses a 163,072-token vocabulary and was trained on more than two trillion tokens of Indic data.

The company claims its tokenizer requires roughly 20% fewer tokens per Indian-language word than the tokenizer used by the GPT-5 family, while using less than half as many as certain byte-level tokenizers used in models such as DeepSeek, Llama and Qwen. Those are company-reported comparisons and should be interpreted as such, rather than as independently audited measurements.

But the underlying idea is technically important regardless of the exact percentage.

Better tokenisation can make AI cheaper.

And in a country of India's scale, cheaper AI changes everything.

Evon v3.3 Has 30 Billion Parameters. But That's Not the Whole Story.

AI companies love talking about parameter counts.

30 billion.

70 billion.

175 billion.

Hundreds of billions.

Soon, trillions.

But parameter count can be misleading.

Gnani's Evon v3.3 has approximately 30 billion total parameters, but only around 3.5 billion parameters are active per token, according to the company's model documentation.

That's because the model uses a hybrid architecture described as a Mamba2-Transformer Hybrid Mixture of Experts, or MoE.

Here's the simplified version.

Instead of activating the entire model every time it processes a piece of text, an MoE architecture can route work through selected parts of the network.

Think of it less like waking up every employee in a company to answer every question.

More like sending the question to the department that is actually needed.

That distinction matters.

A model can have a large total capacity without necessarily requiring every parameter to be used for every computation.

The result can be a much better balance between capability and inference cost.

Gnani says Evon v3.3 can run on a single node and supports a context window of up to 131,072 tokens, commonly described as 128K context. The model card also lists BF16 precision and the hybrid Mamba2-Transformer MoE architecture.

This is where the story starts moving away from the simple question of:

How smart is the model?

And toward a much more useful one:

How economically can you deploy the model?

For enterprises, governments and banks, that may be the more important question.

India Doesn't Have an AI Problem. It Has an AI Economics Problem.

Let's be honest about the challenge.

India has over a billion people.

It has dozens of major languages.

Millions of businesses.

Massive public infrastructure.

Banks processing enormous volumes of customer interactions.

Government systems serving people across wildly different linguistic and economic environments.

An AI system that costs a few dollars per complex interaction may work perfectly well for a premium enterprise in San Francisco.

That doesn't automatically mean it works for India.

India needs AI that can operate at scale without turning every conversation into an expensive GPU bill.

This is why the tokenizer story matters.

This is why sparse architectures matter.

This is why self-hosting matters.

And this is why Gnani Artha is more interesting than a typical "Indian company launches an LLM" announcement.

The company is explicitly positioning the system around what it calls AI economics that work in India.

That phrase might sound like marketing.

But underneath it is a real engineering problem.

The best AI isn't necessarily the smartest AI in the world.

Sometimes it's the smartest AI you can actually afford to run.

11 Indian Languages, Trained Natively

Evon v3.3 supports:

• English
• Hindi
• Bengali
• Telugu
• Tamil
• Marathi
• Gujarati
• Kannada
• Malayalam
• Odia
• Punjabi

Gnani describes these languages as being trained natively on Indic data rather than simply adapting an English-first model afterwards.

That distinction is important.

For years, a lot of multilingual technology followed a predictable pattern:

Build primarily for English. Expand later.

Indian languages were often added through:

• translation layers
• fine-tuning
• additional datasets
• multilingual adapters

That approach can work.

But there is a difference between teaching a system to handle a language and designing the system around that language from the beginning.

India's next generation of AI companies appears increasingly interested in the second approach.

Not:

How do we translate an American AI into Indian languages?

But:

What would AI look like if Indian languages were part of the architecture from day one?

That's a much bigger question.

The Benchmark Numbers Are Impressive. But There's an Important Caveat.

Gnani has published strong benchmark claims for Evon v3.3.

According to its official announcement and model documentation, Evon v3.3 performed strongly on the MILU benchmark, which evaluates Indian-language understanding across multiple languages and subjects.

Gnani says the model outperformed a 105-billion-parameter Indic model on 10 of the 11 languages tested and outperformed a similarly sized 30B model across all 11 languages. The company also reports competitive results against hosted global models on selected evaluations.

Those results are interesting.

But this is where we need to separate excitement from evidence.

The benchmark figures currently come from Gnani's own published testing and model documentation. They should not automatically be treated as independently replicated third-party results.

That doesn't mean they're wrong.

It means the responsible way to report them is exactly what they are:

Company-reported benchmark results.

Independent evaluations over time will tell us much more about where Evon genuinely sits in the global AI landscape.

And honestly, that's fine.

A new model doesn't need to beat every frontier model on every benchmark to matter.

Because Gnani Artha Isn't Just a Model

This is another part of the announcement that people could easily miss.

Gnani Artha is a stack.

Evon v3.3 provides the language intelligence.

But the second component, Plexus, is designed as an agentic AI platform that connects that intelligence to real-world institutional workflows.

This reflects a major shift happening across the AI industry.

The first phase was:

Ask an AI a question.

The next phase is:

Give an AI access to systems and let it perform work.

That could mean interacting with:

• internal databases
• customer systems
• documents
• enterprise software
• APIs
• workflows

The language model becomes the reasoning engine.

The agentic layer becomes the system that turns reasoning into action.

That's where AI stops being a chatbot and starts becoming infrastructure.

And infrastructure is where sovereignty becomes a much bigger conversation.

What Does "Sovereign AI" Actually Mean?

The phrase gets thrown around constantly.

Sovereign AI.

National AI.

Indigenous AI.

Domestic AI.

But what does it actually mean?

In the context of Gnani Artha, a major part of the argument is control over deployment and data.

Because Evon v3.3 is released as open weights, Gnani says organisations can deploy it inside their own data centres or virtual private clouds rather than sending sensitive information to an external public AI service.

That matters enormously for organisations dealing with sensitive data.

Think about:

Banks

Customer financial information.

Insurance companies

Claims, health-related information and personal records.

Government departments

Citizen data.

Healthcare organisations

Patient information.

For these organisations, the question isn't simply:

Which AI model is smartest?

It's also:

Where does the data go?

Who controls the infrastructure?

Can the system operate inside our environment?

Can we keep sensitive information within our own security boundaries?

This is why self-hostable AI may become one of the most commercially important categories in enterprise technology.

And it may be especially important in India.

The Most Interesting Part of Gnani's Strategy Isn't the 30 Billion Parameters

It's this:

Gnani is attacking the "language tax."

Every country has different constraints.

America can afford to throw absurd amounts of capital and computing power at AI.

China has a massive domestic technology ecosystem and enormous state-backed investment.

India has a different advantage.

Scale combined with extreme pressure on cost.

Indian companies have spent decades learning how to serve enormous populations with products that have to be affordable.

That same mindset could become an advantage in AI.

Imagine two models.

Model A

Slightly smarter.

Extremely expensive.

Optimised primarily for global English-language use.

Model B

Slightly less capable on some frontier benchmarks.

Far cheaper to operate in Indian languages.

Faster.

Self-hostable.

Designed for Indian scripts.

Easier to deploy inside regulated institutions.

Which one wins in India?

The answer isn't obvious.

And that's exactly why this story matters.

The AI race may not have one universal winner.

Different markets may reward different kinds of intelligence.

Silicon Valley Is Chasing the Smartest AI. India May Need the Most Useful AI.

This could be India's real opportunity.

Trying to beat OpenAI, Google, Anthropic and the biggest Chinese AI labs in a pure "who has the highest benchmark score?" competition would be brutally expensive.

Those companies operate with extraordinary amounts of capital and compute.

India doesn't necessarily need to copy that strategy.

A different path exists.

Build AI that is:

• multilingual by design
• efficient to serve
• affordable at massive scale
• deployable locally
• useful for regulated industries
• capable of working with Indian languages and scripts without treating them as an afterthought

That's a very different definition of success.

And Gnani Artha is one of the clearest examples yet of an Indian company publicly making that bet.

But There Is One Massive Problem Nobody Can Ignore

Even if India builds brilliant AI models, there is still a deeper dependency hiding underneath the entire AI stack.

Compute.

Training and serving advanced AI models requires enormous amounts of specialised hardware.

GPUs.

Networking.

Data centres.

Electricity.

Cooling.

Semiconductors.

This is where the idea of AI sovereignty becomes complicated.

You can own the model weights.

You can train on domestic data.

You can deploy inside Indian data centres.

You can build Indian software.

But if the underlying hardware stack depends heavily on foreign technology, how sovereign is the system really?

That's not a criticism of Gnani.

It's a question facing virtually every country trying to build national AI capabilities.

The AI stack has layers:

Applications

↓

Agents

↓

Models

↓

Training infrastructure

↓

GPUs and accelerators

↓

Semiconductors

↓

Manufacturing equipment

↓

Global supply chains

A country can become sovereign at one layer while remaining dependent at another.

True technological independence is much harder than launching a domestic chatbot.

And That May Be the Biggest Lesson From Gnani Artha

The future of Indian AI probably won't be decided by one company.

It won't be decided by one model either.

It will depend on whether India can build capability across the entire stack.

Models.

Data.

Talent.

Infrastructure.

Compute.

Semiconductors.

Applications.

Research.

Deployment.

Companies such as Gnani are tackling one important piece of that puzzle.

And their approach is interesting because it doesn't simply copy the Silicon Valley formula.

It starts with a different assumption.

India's languages are not an edge case.

They are the core problem.

The Next AI Revolution in India Might Not Speak English First

For a long time, the internet had an English-first architecture.

Then came smartphones.

Suddenly, hundreds of millions of people came online without needing to use a desktop computer.

India skipped stages of technological development.

Mobile payments exploded.

UPI changed how money moved.

Cheap mobile data changed how people consumed the internet.

AI could follow a similar path.

The next billion AI interactions might not happen through someone typing polished English prompts into a chatbot.

They could happen through:

• Hindi voice conversations
• Malayalam customer service
• Tamil government services
• Kannada business workflows
• Telugu education systems
• Bengali digital assistants

And the companies that win those markets may not be the ones with the largest models.

They may be the ones that understand the languages best.

And can afford to serve them.

Gnani Artha Is a Bet on That Future

It's still early.

The company's benchmark claims will need continued scrutiny.

The global frontier models will keep improving at a frightening pace.

And building a truly sovereign AI ecosystem remains far more difficult than building a single model.

But the technical direction behind Gnani Artha deserves attention.

Because this isn't simply another attempt to create an Indian version of ChatGPT.

The deeper idea is more ambitious.

What if AI built for India shouldn't look like AI built for Silicon Valley at all?

What if the winning model isn't the one with the most parameters?

What if it's the one that processes Indian languages more efficiently?

The one that costs less to run?

The one that can live inside a bank's infrastructure?

The one that can serve millions of people without requiring every interaction to pass through an overseas API?

That is the bet Gnani is making.

And if that bet works, the most important AI innovation coming out of India may not be a bigger model.

It may be a smarter way to build one.

India's AI future might not be about catching up.

It might be about building for a completely different set of rules.