Update: The Evon v3.3 model card on Hugging Face lists an Apache 2.0 licence, 30 billion total parameters with about 3.5 billion active per token, and an August 2026 release. It names 11 supported languages: English plus 10 Indian languages (Hindi, Bengali, Telugu, Tamil, Marathi, Gujarati, Kannada, Malayalam, Odia and Punjabi), so we have corrected an earlier description of “11 Indian languages”. Benchmark scores on the card, such as a MILU macro accuracy of 78.74, are self-reported by Gnani.ai.
New Delhi, August 29: Indian conversational AI company Gnani.ai has launched a sovereign AI stack called Artha. It combines Evon v3.3, an open-weights language model, with Plexus, an agentic platform, and the company is pitching it as an option built around Indian languages, local control and enterprise-grade on-premise use.
Artha was launched on August 28 at the Vice-President’s residence in New Delhi in the presence of Vice-President C. P. Radhakrishnan, according to the Press Information Bureau. Its main offer is that banks, government departments, telecom companies and other institutions can run AI models and agents on infrastructure they control, which could let them use automation and generative AI without sending sensitive data to an outside public cloud.
A new AI product should not be judged on launch claims alone. Many figures about Evon v3.3’s size, speed, token efficiency and language performance come from the company, and public independent benchmarks are still limited. This article therefore keeps confirmed launch details separate from the company’s performance claims.
Key highlights
- Evon v3.3: a 30-billion-parameter open-weights language model that, according to the company, uses a Mixture-of-Experts architecture.
- About 3.5 billion active parameters: the company says only a selected part of the model is active during inference.
- Support for 11 languages, including Indian languages, for building voice and text applications in local contexts.
- Plexus: a platform for building, running and monitoring AI agents that complete workflows.
- Deployment on-premise or in a private environment, so an institution keeps more control over data, access and audit.
What the stack contains
| Component | What it is | Status of the claim |
|---|---|---|
| Evon v3.3 | 30B-parameter open-weights language model, Mixture-of-Experts design | Launch confirmed; performance figures are company-reported |
| Tokenizer | Built for Indian scripts; about 20% fewer tokens than a GPT-5-family tokenizer on Indian words | Company claim; needs independent testing |
| Plexus | Platform to build, run and monitor enterprise AI agents | Launch confirmed; real-world reliability untested |
| Deployment | On-premise or private environment, single-node setup possible | Company description; costs vary by deployment |
Our calculation: if about 3.5 billion of Evon v3.3’s 30 billion parameters are active for each token, roughly 12% of the model (3.5 ÷ 30 ≈ 0.117) does the work at any moment. That is the core reason a Mixture-of-Experts model can be cheaper to run than a dense model of the same total size, though memory still has to hold all 30 billion parameters.
What sovereign AI means
Sovereign AI describes a set-up in which a country or institution has enough control over the model, the computing infrastructure, the training or fine-tuning data and the deployment policies. It is not just a “made in the country” label. The real questions are where the data sits, who controls the model weights, who can audit the system and who decides whether a critical service can be switched off.
In a multilingual country like India, sovereignty also has a language dimension. Most large global models are built on English-centred data and tokenization, and a single word in an Indian script can break into many tokens, making inference slower or costlier. Gnani says Evon v3.3’s tokenizer uses about 20% fewer tokens for Indian words than a GPT-5-family tokenizer. That is a company claim; it can only be confirmed broadly after independent tests across languages and real enterprise workloads.
How Evon v3.3 claims to be different
In a Mixture-of-Experts (MoE) model, the whole model is not equally active for every input; a routing system picks a few expert components as needed. In principle this lets an organisation use the capacity of a large model with relatively little active compute. Gnani says Evon v3.3 can run on a single-node setup, which matters to organisations that do not want to depend on hyperscale cloud because of cost, latency or data policy. Still, hardware specification, throughput, power cost, response quality and the maintenance team needed will differ for each deployment, and “single node” does not automatically mean a lower total cost.
The company says the weights are available under Apache 2.0 licensing. Potential users should read the licence terms, model card, training-data disclosure, safety evaluation and commercial support conditions. Open weights give users more room to inspect and customise, but also more responsibility for security patching and responsible deployment.
What the Plexus agentic platform does
A language model writes answers, while an agentic platform can connect to tools, databases and business rules to complete a goal. Plexus is presented as the layer for building enterprise agents that can handle several steps in customer support, document processing, collections, compliance checks or internal workflows.
A bank’s agent, for example, could understand a customer’s request in their own language, pull limited information from an authorised database, check a policy rule and prepare the next step for a human officer. That is safe only if permissions are minimal, every action is logged and a human reviews high-risk decisions. Giving an AI agent unrestricted system access creates serious operational and privacy risks.
Possible uses in India
Voice interfaces in Indian languages are among Artha’s clearest opportunities. In rural banking, insurance, health appointments, government helplines and farm advisory services, many users do not want to rely on an English keyboard, and combining speech-to-text, a language model and text-to-speech can deliver services in local languages.
On-premise deployment may appeal to regulated sectors, because financial, health or citizen data can stay inside the institution. But data localisation does not by itself guarantee privacy; strong encryption, identity controls, logging, data minimisation and regular security audits are still needed.
Open questions
The first is independent evaluation. Company benchmarks often rest on chosen tasks, hardware and prompts. Judging real quality needs third-party testing across dialects, code-mixed speech, factual accuracy, hallucination rates and long-context reliability in all the supported languages.
The second is training data and safety. Online data in Indian languages is uneven and can carry social biases. The clearer the model card is about data sources, filtering, known limitations and red-team results, the better enterprise buyers can assess risk.
The third is total cost of ownership. Local deployment can cut cloud bills, but the institution then pays for GPUs, power, cooling, engineering, monitoring and upgrades, so the decision should not rest only on the convenience of downloading a model. Transparency also matters for AI regulation and government procurement; our report on the Pentagon–Anthropic court ruling shows that advanced AI is also a question of law and accountability.
What to watch
- Adoption by banks, telecom companies and government departments, and how long deployments run reliably.
- Independent benchmarks and reproducible tests that compare Evon v3.3 with global and Indian alternatives.
- Whether the company publishes fuller model documentation and safety evaluations.
For now, what is confirmed is that the stack has launched, that it includes Evon v3.3 and Plexus, and that the company highlights Indian languages and private deployment as its main features. Claims about speed, token savings and better performance are company-reported and need independent validation.
Frequently asked questions
What is Gnani Artha?
It is an enterprise AI stack made up of the Evon v3.3 language model and the Plexus agentic platform, offered for local control and Indian-language use.
Are all the company’s benchmarks independently verified?
No. Several performance and token-efficiency figures are the company’s own claims, and wider independent testing is still needed.
This article was first published in Hindi and has been translated. The Press of Asia now publishes in English only.
Sources
- Gnani.ai: Artha product and model details
- The Indian Express: Artha launch and technical overview
- The Economic Times: Evon v3.3 and enterprise deployment
- Press Information Bureau: launch event
- Hugging Face: Evon v3.3 model card
The Tech Desk covers technology, artificial intelligence and cybersecurity. Stories are checked against at least two credible sources before publication.
