Indian Startup Gnani AI Unveils Sovereign AI Stack Built Around 30-Billion-Parameter Evon 3.3

India’s push to develop domestically controlled artificial-intelligence infrastructure has gained another major platform with Bengaluru-based Gnani AI launching “Gnani Artha,” an end-to-end sovereign AI stack built around the 30-billion-parameter Evon 3.3 large language model and the Plexus agentic-AI platform.

The system was formally launched by Vice President C. P. Radhakrishnan at Uprashtrapati Bhavan in New Delhi on 28 August 2026. The Vice President’s Secretariat described Evon 3.3 as a large language model and Plexus as a platform designed to connect artificial intelligence with real-world work and institutional processes.

Artha is intended primarily for Indian enterprises and public institutions that require AI systems capable of handling Indian languages while allowing sensitive information and model workloads to remain within their own computing infrastructure.

That combination of Indic-language capability, open model weights, self-hosted deployment and enterprise AI agents gives Artha significance beyond the launch of another chatbot or language model. It represents an attempt to build more of the AI technology stack under Indian control.

Artha Combines AI Models, Agents and Deployment Infrastructure

Gnani describes Artha as an end-to-end sovereign AI solution integrating three broad layers: AI models, an agentic orchestration platform and the infrastructure on which they are deployed.

At the centre is Evon 3.3, a 30-billion-parameter language and reasoning model designed for Indian-language workloads.

The second major component is Plexus, Gnani’s agentic AI platform. Rather than simply answering questions, Plexus is designed to create and coordinate AI agents capable of carrying out multi-stage tasks across documents, databases, enterprise applications and conversations.

The third element is deployment. Gnani says organisations can operate Artha inside their own data centre or virtual private cloud, allowing customer information to remain within the organisation’s infrastructure rather than being routinely sent to an external AI provider.

For banks, insurers, public-sector bodies and other organisations handling regulated or sensitive information, this deployment model is an important part of the system’s sovereign-AI proposition.

Evon 3.3 Has 30 Billion Parameters but Activates Around 3.5 Billion at a Time

Evon 3.3 is technically notable for its mixture-of-experts architecture.

Although the model contains approximately 30 billion total parameters, only around 3.5 billion parameters are activated for a particular token or computation, according to Gnani.

This architecture can reduce the computational resources required during inference. Instead of activating the entire model for every operation, different groups of parameters specialise in different types of tasks and only a subset is used at a given time.

Gnani says this allows Evon 3.3 to operate on a single inference node, rather than requiring a large multi-node cluster. The company lists the system as a 30B/3.5B-active model and says it can be self-hosted.

The Economic Times reported that Gnani has tested deployment on hardware including Nvidia RTX 6000 Pro and L40S-class accelerators, potentially allowing enterprise deployment without depending exclusively on the most expensive frontier AI processors.

Built Around Indian Languages Rather Than Translation Alone

One of Evon 3.3’s central design objectives is addressing the persistent imbalance between English and Indian languages in generative AI.

Gnani says the model has been trained across 11 Indian languages using Indic-language data, rather than depending entirely on an English-first model followed by translation.

According to company executives interviewed by The Economic Times, the training corpus contained more than two trillion tokens across 11 Indian languages.

The model has also been optimised for language understanding, reasoning, tool use, response speed and computational cost.

This is especially relevant in India because real-world institutional data rarely exists only in English.

Loan applications, government forms, citizen grievances, healthcare records, customer-service conversations and legal or administrative material may combine English with Hindi, Kannada, Tamil, Telugu, Marathi and numerous other Indian languages.

India therefore requires AI systems capable not simply of translating words but of reasoning over documents and conversations produced within India’s multilingual environment.

Gnani Rebuilt the Tokenizer for Indian Scripts

One of the more technically interesting features of Evon 3.3 is its tokenizer.

Before a language model processes a sentence, the text is divided into smaller computational units known as tokens. The number of tokens required to represent a word or sentence directly affects inference cost, processing time and the amount of useful information that can fit within the model’s context window.

Tokenizers developed primarily around English can represent some Indian scripts inefficiently. A word that requires only a small number of tokens in English may require substantially more computational units when written in an Indic script.

Gnani says it rebuilt Evon 3.3’s tokenizer specifically to address this problem.

The company claims its tokenizer requires approximately 20% fewer tokens per Indian-language word than the tokenizer used by the GPT-5 family, while also substantially reducing token consumption compared with byte-level tokenisation approaches.

If replicated across millions of documents or conversations, even relatively modest token reductions can have substantial consequences for operating cost and latency.

This is particularly important for Indian banks, telecom operators, government services and customer-service platforms that may process millions of multilingual interactions.

Open Weights Under Apache 2.0 Licence

Evon 3.3 is being offered as an open-weight model.

Gnani says the model weights are available by request through Hugging Face under the Apache 2.0 licence, allowing organisations to deploy and adapt the model rather than being restricted entirely to a closed external API.

The distinction between open weights and a closed cloud model is strategically important.

When an organisation uses a proprietary AI service hosted abroad, it may have limited control over the underlying model, deployment environment, future modifications and long-term pricing.

With open weights, an enterprise can operate the model within its own infrastructure, fine-tune it for specialised requirements and retain greater control over the deployment architecture.

That is one of the central ideas behind the increasingly important concept of sovereign AI.

What Sovereign AI Means

Sovereign AI does not necessarily mean that every semiconductor, software library or foundational technology used by a system must originate within national borders.

Rather, the concept generally concerns the ability of a country or institution to maintain meaningful control over its models, data, computing infrastructure, deployment decisions and strategic AI capabilities.

For Artha, sovereignty is principally expressed through self-hosting, Indian-language optimisation, open model weights and the ability for institutions to keep sensitive information within their own infrastructure.

Gnani says Artha can run inside an organisation’s own data centre or virtual private cloud, meaning customer information does not need to leave that network for external model inference.

Such an architecture could be particularly relevant for sectors governed by strict rules concerning privacy, financial information and data residency.

Evon 3.3 Has a Technical Link to Nvidia’s Nemotron

The description of Evon 3.3 as an indigenous Indian AI model requires some technical context.

According to detailed reporting by The Economic Times, Gnani developed Evon 3.3 through continual pre-training of an Nvidia Nemotron model using Gnani’s own Indic-language and domain-specific datasets, followed by post-training and reinforcement learning.

Gnani executives said the company introduced its own training data, tokens, language optimisation and post-training techniques to substantially alter the resulting model’s capabilities.

The company therefore controls and develops the Evon system and has substantially customised its architecture for Indian workloads, but the published technical reporting indicates that its lineage includes Nvidia’s Nemotron technology.

This distinction is worth retaining when describing the system. Artha is an Indian-developed sovereign AI stack, while Evon 3.3’s development reportedly incorporates an existing Nvidia model foundation that Gnani has extensively retrained and optimised.

Around 1,500 GPUs Used During Development

Developing modern foundation models requires substantial computing infrastructure.

Gnani co-founder and Chief Product and Engineering Officer Bharath Shankar told The Economic Times that the development process involved data cleaning, continual pre-training and post-training, with around 1,500 Nvidia GPUs used across the different stages.

The figure illustrates one of the fundamental challenges India faces in building a domestic AI ecosystem.

Advanced AI models require not only algorithms and datasets but also access to large quantities of high-performance compute capacity.

This is one reason the Government of India has made shared computing infrastructure a major component of the IndiaAI Mission.

Gnani Is Part of India’s Foundation-Model Programme

Gnani AI is not operating in isolation from India’s wider sovereign-AI programme.

The company is among 12 organisations and consortia selected under the IndiaAI Mission to develop indigenous foundational models and large or small language models based on Indian datasets.

The Government of India confirmed in February 2026 that the selected organisations include Gnani AI, Sarvam AI, Soket AI, Gan AI, BharatGen through an IIT Bombay consortium, Tech Mahindra Maker’s Lab and several other Indian AI developers.

The IndiaAI Mission was launched with an approved outlay of approximately ₹10,372 crore and includes programmes for computing infrastructure, foundation models, datasets, applications, startup financing, skills and safe AI development.

By February 2026, the government said more than 38,000 GPUs had been onboarded for the common compute facility.

Government support for domestic models is intended to reduce one of the biggest barriers facing Indian AI startups: the enormous cost of computing required for large-scale training and experimentation.

Gnani’s Earlier IndiaAI Work Focused on Voice AI

Gnani already had an important role within the IndiaAI programme before the launch of Artha.

Government documentation previously identified Gnani AI as developing a 14-billion-parameter Voice AI foundation model intended to provide multilingual real-time speech processing and reasoning capabilities.

The company has strong roots in conversational and voice artificial intelligence, including speech recognition, speech synthesis, voice biometrics and customer-service automation.

Artha significantly broadens that technological footprint.

The new stack incorporates not merely voice processing but language reasoning, enterprise agents, workflow orchestration and self-hosted foundation-model deployment.

Artha Includes More Than Evon 3.3

Although Evon 3.3 is the headline model, Gnani describes Artha as incorporating a broader family of AI technologies.

These include Evon v2.0, intended for enterprise reasoning and tool orchestration; Prisma v2.5, an automatic speech-recognition system designed for Indian accents and code-switching; and Timbre v2.5, a neural speech-synthesis model.

These systems complement Evon 3.3 by allowing Artha to work not only with written language but also with spoken interactions.

The combination is particularly relevant in India, where many citizens may interact with digital services more naturally through speech than through conventional text interfaces.

Plexus Turns Language Models Into AI Workers

The second major pillar of Artha is Gnani Plexus.

A large language model by itself primarily generates or analyses information. An agentic platform attempts to go further by allowing AI systems to take actions, call software tools and complete sequences of tasks.

Gnani describes each Plexus agent as a discrete unit that can be combined with other agents into workflows centred around a particular outcome.

Those workflows can operate sequentially or in parallel, interact with existing enterprise systems and incorporate either human approval or AI-led orchestration.

This means that instead of merely asking an AI model to explain a bank transaction, an enterprise could theoretically build an agent that retrieves records, checks them against another system, identifies discrepancies, prepares a resolution and sends the case to a human when required.

Potential Applications in Banking and Insurance

Financial services represent one of the most obvious markets for Artha.

India’s banking and insurance industries handle enormous numbers of documents, telephone conversations, compliance checks and transactions, frequently across several languages.

Gnani lists potential applications including loan-document processing, underwriting, bank reconciliation, KYC remediation, transaction-dispute management and customer grievance resolution.

For example, an agent could examine a loan application alongside bank statements, GST documents and identity records, compare the information against authorised databases and refer exceptional cases to human staff.

Another workflow could reconcile millions of transactions across different financial systems and identify records requiring investigation.

These examples illustrate the type of operational automation Plexus is intended to support.

They should not, however, be interpreted as evidence that these specific systems are already deployed by particular banks or government departments. Gnani itself explicitly describes the use cases as illustrative rather than existing customer deployments.

Government Services Are Another Major Target

The public sector could present an equally important application area.

Gnani proposes workflows involving multilingual citizen grievance management, welfare-scheme enrolment, failed benefit-transfer resolution, public-health outreach and disaster-response coordination.

A citizen could theoretically communicate a grievance in his or her own language, after which one AI agent records it, another identifies similar complaints, another routes the issue to the appropriate department and the system subsequently informs the citizen when action has been completed.

Similar architectures could be used to identify eligible citizens who have not enrolled in welfare schemes or to detect failed benefit transfers caused by account mismatches.

Such systems could potentially reduce administrative workloads, although real-world government adoption would require strong safeguards covering accuracy, privacy, auditability, cybersecurity and human oversight.

Guardrails and Auditability Become Crucial With Agentic AI

The ability of AI systems to take actions rather than merely generate text creates a different class of risk.

An inaccurate chatbot response is problematic. An inaccurate autonomous system capable of changing financial records, sending payments or making administrative decisions can be considerably more consequential.

Gnani says Plexus incorporates guardrails, observability, audit logging, prompt-defence mechanisms, jailbreak resistance and data-protection controls.

Human intervention can also be incorporated at designated stages of a workflow.

These controls will be crucial if agentic systems eventually move into high-stakes areas such as banking, insurance, healthcare or government administration.

Gnani Reports Strong Indian-Language Benchmark Performance

Gnani has published ambitious performance claims for Evon 3.3.

On MILU, an Indian-language benchmark covering multiple academic and professional subjects across 11 languages, the company says Evon 3.3 outperformed an unnamed 105-billion-parameter Indic model in 10 of the 11 languages evaluated.

Gnani also says it surpassed another similarly sized 30-billion-parameter model across all 11 languages and achieved performance comparable with a similar-sized hosted global model.

The Economic Times reported that the results were based on Gnani’s testing across approximately 40–45 benchmarks, including MILU, MMLU and MMLU-Pro.

These results are promising, but they should presently be treated as company-reported benchmark results rather than universally established independent rankings.

Independent evaluations and wider developer testing will provide a clearer picture of Evon 3.3’s capabilities as the model becomes more broadly available.

Smaller Active Model Could Lower AI Costs

Raw parameter count is not necessarily the most important measure of an AI system.

For enterprise adoption, cost per interaction can become equally important.

Evon 3.3’s mixture-of-experts architecture means approximately 3.5 billion parameters are active at a time despite the model’s total 30-billion-parameter size.

Combined with its Indic-optimised tokenizer, Gnani argues that this makes the system considerably cheaper to operate on Indian-language workloads.

The company has claimed that Evon 3.3 could cost between roughly one-third and one-fifth as much as certain comparable proprietary AI models, depending on deployment and workload.

Again, these are company estimates rather than independently established universal cost comparisons.

Actual economics will depend on hardware, throughput, model configuration, context length and the type of application being run.

Why Indian-Language AI Economics Matter

The cost issue is particularly significant in India because the country’s potential AI user base is enormous.

A model serving a relatively small group of highly paid professionals can tolerate expensive inference. A system processing government interactions, banking calls or healthcare communication for tens of millions of people requires a different economic model.

If Indian-language text consumes significantly more tokens than English, the cost disadvantage becomes magnified at national scale.

Reducing that so-called “language tax” could therefore determine whether advanced AI becomes practical for mass-market services in Indian languages.

It is one of the more important aspects of Evon 3.3 because it addresses not simply language quality but the economics of deploying AI across India.

AI Sovereignty Is Becoming a Strategic Issue

The launch of Artha comes as artificial intelligence increasingly becomes part of national technological strategy.

A small number of global companies currently dominate frontier AI models, accelerator hardware and hyperscale cloud infrastructure.

Heavy dependence on these systems can expose countries and institutions to foreign pricing decisions, technology restrictions, data-governance concerns and changes in access to proprietary models.

India’s sovereign-AI strategy therefore seeks to develop domestic capabilities across several layers: compute infrastructure, datasets, foundation models, applications and AI talent.

The objective is not necessarily technological isolation.

Rather, it is to ensure that India retains the ability to develop and operate important AI systems even while participating in the global technology ecosystem.

Vice President Emphasises Moving From Consumer to Creator

Speaking at the launch, Vice President C. P. Radhakrishnan linked Artha to the broader objective of technological self-reliance.

He argued that India should not remain perpetually dependent on technologies created elsewhere and said the country’s engineers must increasingly become creators of frontier technology.

The Vice President also pointed to the IndiaAI Mission and Centres of Excellence for AI as part of India’s expanding artificial-intelligence ecosystem and described sovereign AI capability as important to the country’s technological and strategic autonomy.

His remarks reflect a wider shift in India’s technology policy from adoption towards domestic creation of foundational platforms.

From Software Services to Foundation Models

India has long been a major global centre for software development, IT services and engineering talent.

Foundation-model development represents a different challenge.

Building sophisticated AI models requires large datasets, specialist researchers, high-performance processors, model-training expertise and considerable financial resources.

The emergence of companies developing Indian foundation models therefore represents an attempt to move further up the technology stack—from creating applications on top of foreign platforms towards owning more of the underlying intelligence layer itself.

Gnani’s Artha launch is one manifestation of that transition.

The Important Test Comes After the Launch

Artha’s specifications are significant, but the true measure of the platform will emerge through deployment.

Developers and enterprises will need to evaluate how well Evon 3.3 performs outside controlled benchmarks, particularly on difficult reasoning tasks, Indian-language code-switching, hallucination rates and domain-specific workloads.

Organisations will also need to determine whether Plexus can reliably coordinate agents across complex enterprise systems without creating unacceptable operational or security risks.

Independent benchmarking will be particularly valuable in assessing Gnani’s performance and cost claims.

AI models frequently perform differently in real applications than they do on standardised evaluation datasets.

The availability of open weights should make broader external testing easier.

An Important Addition to India’s Sovereign AI Ecosystem

Gnani Artha is significant not because India suddenly possesses a single AI system capable of replacing the world’s largest frontier models.

Its importance lies elsewhere.

Artha combines a 30-billion-parameter Indian-language model, an agentic workflow platform, open weights and self-hosted deployment into a technology stack directed specifically at Indian institutional requirements.

Evon 3.3 attempts to address one of the less visible barriers to India’s AI adoption: the computational penalty associated with processing Indian scripts. Plexus attempts to turn the resulting intelligence into operational workflows. Self-hosted deployment addresses concerns surrounding control over sensitive data.

Together, these capabilities provide another building block for India’s emerging sovereign-AI ecosystem.

The platform also demonstrates how India’s AI strategy is evolving from conversational assistants and individual language models towards complete AI stacks capable of being deployed inside banks, enterprises and public institutions.

Nevertheless, the launch of Gnani Artha on 28 August 2026 marks another step in India’s effort to shift from being principally a consumer of global artificial-intelligence technology towards becoming a developer of AI models, platforms and infrastructure designed around its own languages, institutions and strategic requirements.

---

References

1. Press Information Bureau, Vice President’s Secretariat — “Vice President Shri C. P. Radhakrishnan launches ‘Gnani Artha’, a sovereign AI stack comprising Evon 3.3 and Plexus,” 28 August 2026. Official confirmation of the Artha launch at Uprashtrapati Bhavan and its two principal components.
2. Gnani AI — “Gnani Artha: The Next Frontier of Sovereign AI.” Official technical information covering Evon 3.3, its 30B/3.5B-active architecture, 11-language support, tokenizer optimisation, Apache 2.0 open weights, Plexus and self-hosted deployment.
3. The Economic Times — “Gnani AI launches Artha sovereign AI stack with 30-billion-parameter Evon 3.3,” 28 August 2026. Provides technical details on the Nemotron lineage, two-trillion-token training corpus, mixture-of-experts architecture, GPU usage, token efficiency and benchmark claims.
4. The Indian Express — “Gnani unveils sovereign AI stack ‘Artha’ featuring open-weight model, enterprise agents,” 28 August 2026. Details enterprise use cases, Plexus agentic capabilities, self-hosting and Gnani’s place within India’s sovereign-AI programme.
5. Press Information Bureau, Ministry of Electronics & IT — “In less than 24 months, India AI Mission has Set up a Foundation for Development of AI Ecosystem in the Country,” 13 February 2026. Confirms Gnani AI among 12 organisations selected for indigenous foundation-model development and outlines IndiaAI compute infrastructure.
6. Press Information Bureau — “Development of India’s Foundational Models.” Details government support for Gnani AI’s multilingual Voice AI work and the broader IndiaAI foundation-model programme.

---