Artha: the Indian sovereign AI stack that cheapens costs with LLM of open weights

Gnani, an Indian artificial intelligence company, announced Artha, a sovereign AI stack aimed at large companies in the country. The solution includes the Evon v3.3 language model, with 30 billion parameters and support for more than 11 Indian languages, and a reconstructed tokenizer that uses about 20% less tokens per Indian word compared to GPT-5. The goal is to allow local organizations to implement generative AI with control over data and costs.

According to the TechGig portal, Artha stands as an alternative to dependence on foreign infrastructure, combining digital sovereignty with operational efficiency for the Indian corporate market.

Sovereign AI: What is Stack Artha?

Artha is a complete platform that brings together the open weight language model, adaptation tools, optimized tokenizer and local deployment infrastructure. The proposal is that companies can run their own AI systems without resorting to international clouds or closed APIs, keeping the data within their borders. The name “Artha” refers to one of the goals of life in Indian philosophy — economic prosperity —, which reinforces the commercial intention of the solution.

The technical highlight is the Evon v3.3, an LLM with 30 billion parameters capable of processing more than 11 Indian languages, including Hindi, Bengali, Tamil and Telugu. By having open weights, the model can be downloaded, adjusted and run on own servers, reducing dependency on external suppliers.

Why does India bet on sovereign AI?

Digital sovereignty has become a strategic priority in several countries, and India is no exception. With a billion-dollar population and an expanding startup ecosystem, the country wants to ensure that its sensitive data is not processed in infrastructure controlled by other governments or corporations. In addition, Indian linguistic diversity requires efficient native models — the reconstructed Evon v3.3 tokenizer was designed exactly to solve this bottleneck, since international models are usually less efficient in languages such as Hindi or Tamil.

The bet on open weights also expands the control of companies over the AI life cycle, allowing audit. This dialogues with a global scenario where AI tools often operate without proper supervision.

How does LLM Evon v3.3 reduce costs?

The cost of a language model is directly related to the number of tokens processed in each query. The Evon v3.3 tokenizer was reconstructed from scratch to represent Indian words with fewer fragments: the reduction is approximately 20% compared to GPT-5. In high-volume applications, this economy translates into smaller computing and inference invoices, as well as lower latency by processing fewer tokens.

This gain is relevant for companies that plan to scale customer service, document analysis and automation in regional languages. An efficient tokenizer can also influence the quality of the response, since the model allocates more attention capacity per original word.

Sovereign AI in Brazil: the impact here

The Gnani initiative shows a viable path for developing countries to build AI layers aligned with their linguistic and regulatory realities. In Brazil, the discussion on digital sovereignty is on the agenda, especially with AI regulation projects and the growing use of open models by local companies. In a sovereign AI analysis, we show how this movement connects to corporate regulation and adoption.

For developers, Artha is another open model option that escapes the big tech domain. However, it is necessary to evaluate the documentation and technical community around Evon v3.3 before adopting it in production, especially for use outside India.

Frequently Asked Questions about Stack Artha

Is the Evon v3.3 model available to any company?

From source information, Gnani initially presented Artha to Indian companies, although the model has open weights — suggesting potential public access, but without official details on licensing outside India.

What does an LLM of open weights mean?

It means that the model's trained parameters are available for download and adaptation by developers. This allows you to run the template on your own servers, adjust it with specific data and integrate it to internal systems without relying on closed APIs.

How does the efficiency of the tokenizer translate into lower cost?

As the cost per query in language models is calculated by the number of processed tokens, a tokenizer that uses 20% less tokens per word proportionally reduces the cost of each call. For those who operate on a scale, this difference generates significant savings in AI infrastructure.