Gnani AI Launches Artha Sovereign AI Stack With 30-Billion-Parameter Evon 3.3 Model

Gnani AI Unveils Artha Sovereign AI Stack for Indian Enterprises

Bengaluru-based voice AI company Gnani AI launched Artha, a sovereign AI stack designed specifically for Indian enterprises and public institutions. The stack was unveiled in New Delhi by Vice President C.P. Radhakrishnan on August 28 at Upa-Rashtrapati Bhavan.1

1

2

2

Artha represents a strategic move toward technological self-reliance, combining the 30-billion-parameter open-weights model Evon 3.3 with Plexus, an agentic platform that enables organisations to build AI agents for specific workflows.3

3

The Artha AI stack addresses three critical challenges facing Indian organisations adopting AI: data control, the cost of scaling AI operations, and effective performance across Indian languages in real-world scenarios.3

3

For banks, insurers, and government departments handling sensitive information, Artha allows organisations to run AI models and applications entirely within their own infrastructure, meeting DPDP, RBI, and IRDAI residency requirements without customer data ever leaving their systems.1

1

2

2

Evon 3.3 Language Model: Built on Nvidia's Nemotron Foundation

The Evon 3.3 language model forms the intelligence core of the Artha AI stack. Gnani AI developed this multilingual language model by continually pre-training Nvidia's Nemotron model on more than 2 trillion tokens across 11 Indian languages, followed by post-training and reinforcement learning.1

1

According to Gnani co-founder and CEO Ganesh Gopalan, the model demonstrates significantly better performance for Indian languages compared to generic global models in terms of tokens consumed and accuracy on benchmarks such as MILU.1

1

Evon 3.3 employs a mixture-of-experts architecture where, despite having 30 billion parameters, only approximately 3.5 billion are activated for any given task.1

1

This design allows the model to deliver reasoning and language-understanding capabilities while consuming less computing power than larger models. Gnani's chief product and engineering officer Bharath Shankar explained that the model has been optimised for tool calling, language understanding, reasoning, speed, and cost.1

1

Performance Benchmarks and Cost Efficiency for Indian Languages

Gnani AI claims that Evon 3.3 outperforms Sarvam AI'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.1

1

The results are based on internal testing across approximately 40-45 benchmarks, including widely used standards like MMLU and MMLU-Pro.1

1

On these benchmarks, Evon 3.3 achieves parity with similar-size hosted global frontier models.2

2

Gnani AI rebuilt the model's tokenizer to handle Indian scripts more efficiently. The company claims Evon 3.3 requires approximately 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.1

1

2

2

This efficiency translates directly into lower inference costs and reduced latency. Running Evon 3.3 could cost between one-third and one-fifth as much as comparable OpenAI models, with roughly 39% to 40% greater efficiency based on tokenizer performance.1

1

Plexus Agentic Platform: From Intelligence to Implementation

The Plexus agentic platform transforms the language intelligence of Evon 3.3 into practical workflows for AI for Indian enterprises and AI for public institutions.2

2

Each agent functions as a discrete, identity-bearing unit that can be combined with other agents into workflows built around defined outcomes. In grievance resolution by government agencies, for example, a multilingual AI agent captures citizen complaints while a reasoning agent detects patterns across recent complaints, files consolidated tickets with responsible departments, and closes the loop with citizens once issues are resolved.2

2

For bank reconciliation, an AI agent matches millions of line items across bank statements, core-banking ledgers, and payment-switch logs, clearing clean matches and drafting root-cause narratives for genuine exceptions.2

2

These workflows run under an orchestration layer that can be human-in-the-loop or AI-led, with guardrails, observability, and audit logging built in. Plexus integrates with enterprise and public digital infrastructure, supports tool calling, and works across documents, systems, and conversations.2

2

Data Sovereignty and On-Premises Deployment Capabilities

The Artha AI stack directly addresses data sovereignty concerns that have become increasingly urgent for Indian organisations. Because Evon v3.3 ships as open weights that run on a single node, it can be deployed entirely inside an institution's own data centre or virtual private cloud.2

2

This on-premises AI deployment model is particularly relevant for regulated industries where customer data cannot leave institutional infrastructure.

The model can run on hardware such as Nvidia's RTX 6000 Pro or L40S and does not require high-end accelerators such as the B200.1

1

Gnani AI used approximately 1,500 Nvidia GPUs across data cleaning, continual pre-training, and post-training stages during development.1

1

Model weights are available upon request through Hugging Face under an Apache 2.0 license, though Gnani AI controls access despite the open licensing.3

3

Early Customer Traction and Use Cases

Adoption of the enterprise AI stack remains in early stages. According to CEO Ganesh Gopalan, the company has begun presenting the model to select customers. At a recent customer meeting in Pune, five of the 20 participating enterprises expressed interest and have started building on Evon.3

3

Potential use cases include underwriting, advertising, and payments reconciliation, though these initiatives have not yet reached deployment stage.3

3

Gopalan noted that while some voice AI workflows are mostly autonomous, regulated applications such as underwriting and payments reconciliation require additional oversight before agents can fully implement them.3

3

The real test will be whether banks, insurers, and public institutions trust these agents with regulated workflows in production environments.

Strategic Context: India's Push for Sovereign AI

The launch of Artha fits within India's broader push for technological self-reliance and Sovereign AI capabilities. Vice President Radhakrishnan stated that combining Evon 3.3 and Plexus demonstrates the increasing strength of India's technology ecosystem and shows that Indian engineers can both use and develop advanced technologies.3

3

He emphasised the need to build sovereign AI capabilities to enhance India's technological and strategic autonomy.3

3

India's push has become more urgent due to concerns about relying on US and Chinese AI infrastructure for sensitive government and enterprise data, as well as the high costs of running large foreign models.3

3

The IndiaAI Mission, launched in 2024, supports several domestic model initiatives by providing subsidised GPU compute. Sarvam AI, a flagship grantee, has released open-source Indian-language models promoting voice-first design, efficient models, agentic capability, and data residency. BharatGen, an academic-industry consortium led by IIT Bombay, has adopted a similar open-weights strategy.3

3