India’s Big Step Towards Self-Reliant AI with Gnani Artha

India has taken another significant step towards building self-reliance in artificial intelligence (AI). Bengaluru-based voice AI startup Gnani.ai has launched its indigenous AI stack, “Gnani Artha,” in New Delhi.

The platform was launched by Vice President C. P. Radhakrishnan and is built around Gnani.ai’s foundational Evon v3.3 model and enterprise AI platform Plexus.

Designed specifically for India, the open-weight AI stack supports 11 Indian languages and aims to offer lower operating costs, faster processing and stronger data security compared with many foreign AI models.

What Makes Evon v3.3 Different?

At the core of Gnani Artha is Evon v3.3, a 30-billion-parameter open-weight AI model trained to handle India’s diverse linguistic environment.

The model is based on a Mixture-of-Experts (MoE) architecture. Although it has 30 billion parameters in total, only around 3.5 billion parameters are activated for each token.

This approach allows the system to use only the computing resources required for a particular task rather than activating the entire model for every query. The result is potentially faster responses, lower power consumption and reduced computing costs.

The model is expected to be made available to developers through Hugging Face under the Apache 2.0 licence, enabling organisations and developers to customise and deploy it for their own applications.

Built for Indian Languages

One of the key features of Evon v3.3 is its focus on Indian languages and scripts.

Gnani.ai has developed a new tokenizer designed around Indian scripts. According to the company, the model uses around 20% fewer tokens per word than the GPT-5 family for Indian-language processing and less than half the token count used by models such as Llama, DeepSeek and Qwen.

Lower token usage can translate into reduced inference costs and faster responses, while also making extended AI conversations more economical.

Competing with Larger AI Models

Gnani.ai says Evon v3.3 has demonstrated strong performance on the MILU-11 benchmark, which evaluates AI capabilities across Indian languages.

According to the company, its 30-billion-parameter model has outperformed larger models, including Sarvam’s 105-billion-parameter model, across the benchmark’s 11 languages.

The company says the model delivers performance that can be competitive with global AI models such as GPT-5.4 Nano in certain evaluations, despite having significantly fewer parameters.

Plexus: Powering AI Agents

The second major component of Gnani Artha is Plexus, an enterprise platform designed to build and deploy intelligent AI agents.

These agents can perform multiple steps in a workflow, allowing them to handle complex tasks that would traditionally require human intervention.

Plexus also incorporates enterprise security and governance features, including audit logging, guardrails and human-in-the-loop controls.

The platform can run on a single node, giving banks, insurance companies and government departments the option to deploy AI within their own data centres or cloud environments. This can help organisations maintain greater control over sensitive customer and operational data.

Gnani.ai, which was founded in 2016, currently works with more than 200 large enterprises and is also part of the Government of India’s IndiaAI Mission.

Where Can Gnani Artha Be Used?

Faster Government Grievance Resolution

Gnani Artha could be used to streamline the handling of public grievances.

Its multilingual AI agents can register complaints in different Indian languages. A reasoning agent can then analyse the complaint and determine whether it relates to a specific district or represents a wider issue.

The system can route the complaint to the appropriate government department and, once the matter is resolved, provide an update directly to the citizen.

Automated Banking Reconciliation

The technology could also have applications in the banking sector, particularly in transaction reconciliation.

AI agents can analyse millions of records by cross-checking information from bank statements, core-banking ledgers and online payment-switch logs.

Mismatched or problematic entries can be identified automatically. Where a transaction requires human intervention, the system can escalate it to a bank official along with a clear explanation of the discrepancy.

A Step Towards India-Centric AI

The launch of Gnani Artha highlights the growing push to develop AI systems specifically suited to India’s languages, data requirements and enterprise needs.

By combining an open-weight foundation model with an enterprise AI-agent platform, Gnani.ai is positioning its technology as a potential alternative for organisations seeking lower-cost, multilingual and locally deployable AI solutions.