On August 28, 2026, Vice President C.P. Radhakrishnan stood at his New Delhi residence and unveiled what could become India's most significant AI milestone yet: Artha, a sovereign AI stack from Bengaluru-based Gnani AI that promises to run Indian enterprise AI entirely on Indian infrastructure, in Indian languages, for Indian costs.
The launch wasn't just another product announcement. It came with a model — Evon v3.3 — that claims to use 20% fewer tokens per Indian-language word than OpenAI's GPT-5 family, and less than half what Chinese open-weight models like DeepSeek, Llama, and Qwen require. For a country where AI adoption has been held back by both cost and language complexity, this isn't incremental improvement. It's a different arithmetic.
The Sovereignty Play
Artha combines two components: Evon v3.3, a 30-billion-parameter open-weight language model, and Plexus, an agentic AI platform for enterprise deployment. Both are designed for self-hosting on a single node — meaning banks, insurers, and government departments can keep sensitive customer data within their own infrastructure, compliant with India's growing data residency requirements.
"Sovereign AI is not about keeping the world out," said Gnani CEO Ganesh Gopalan. "It is about India having the capability to build for itself — and then for every country that shares its problems."
The timing is telling. As China's DeepSeek sent shockwaves through Silicon Valley with its cost-efficient open-weight models, and as the US tightens AI export controls, India is carving its own path: build locally, open globally.
How Evon 3.3 Actually Works
Evon v3.3 is built on Nvidia's Nemotron foundation model, but Gnani didn't stop there. The company continually pre-trained the model on over 2 trillion tokens across 11 Indian languages — Hindi, Tamil, Telugu, Bengali, Marathi, Gujarati, Kannada, Malayalam, Odia, Punjabi, and Urdu — using domain-specific data from Indian enterprise use cases.
The result is a mixture-of-experts architecture where only about 3.5 billion of the 30 billion parameters are activated for any given task. This design allows Evon to deliver reasoning and language capabilities while consuming significantly less compute than larger dense models.
"You do not see those capabilities in the standard Nemotron model," said Gnani co-founder and CPO Bharath Shankar. "But you see them in Evon."
The proof is in the benchmarks. On MILU — a benchmark specifically designed to evaluate Indic language understanding across academic and professional subjects — Evon v3.3 outperformed both a similarly-sized unnamed model and an unnamed 105-billion-parameter Indic model in 10 out of 11 languages.
The Token Economy
Here's where the economics get interesting. Token consumption has become a major cost driver for enterprises adopting AI, especially for multilingual deployments. Gnani rebuilt Evon's tokenizer specifically for Indian scripts, achieving what the company calls "significantly better token fertility" — essentially, more meaning per token.
The numbers: Evon 3.3 requires roughly 20% fewer tokens per Indian-language word than GPT-5's tokenizer, and less than half the tokens used by byte-level tokenizers in models like DeepSeek, Llama, and Qwen. For an enterprise processing millions of customer interactions daily across multiple Indian languages, this translates to real cost savings — not just marketing claims.
This matters because the open-weight model war is shifting from pure capability competition to efficiency competition. DeepSeek proved that Chinese labs could match Western performance at a fraction of the cost. Now Gnani is making the same move for India — but with a focus on linguistic sovereignty rather than just price arbitrage.
Plexus: The Agent Layer
While Evon is the brain, Plexus is the nervous system. The agentic AI platform allows enterprises to build and deploy AI agents through natural language prompts, with support for tool calling and autonomous operation across documents, systems, and conversations.
Gnani demonstrated Plexus with two use cases: an agent that fetches customer PAN card details with a single prompt, and a multi-agent swarm for welfare beneficiary programs and grievance resolution — exactly the kind of government-scale automation India needs.
The platform supports multiple underlying models, including Evon v3.3, giving enterprises flexibility while maintaining the sovereign stack promise.
The Bigger Picture
India's AI trajectory is diverging from both the Western commercial model and the Chinese state-control approach. Instead, it's pursuing what Vice President Radhakrishnan called "open, affordable, and accessible" AI — technology that uplifts society rather than concentrating power.
The launch came at a moment when AI nationalism is accelerating globally. The US is restricting chip exports to China. China is building domestic AI champions. Europe is crafting its own regulatory framework. And India? India is building models that speak its languages, run on its infrastructure, and serve its people.
As Gopalan put it: "Both these platforms reflect the growing strength of India's technology ecosystem. This initiative shows that our engineers have the capability not only to use frontier technologies, but also to build them."
The question now is whether Artha can scale beyond the pilot phase and compete with the well-funded open-weight models emerging from China. The technology looks promising. The timing is right. Whether the market follows remains to be seen.
🔥 Hot Takes
1. India is playing a different game than China. DeepSeek won attention by matching Western performance at rock-bottom prices. Gnani is taking a more nuanced approach: match Western performance, beat Chinese efficiency on Indian languages, and add sovereignty as a third dimension. It's not just about being cheap — it's about being relevant.
2. Token efficiency is the new parameter count. The AI industry obsessed over model size for years. Now the competitive advantage is shifting to whoever can deliver the same quality with fewer tokens. Gnani's 20% improvement over GPT-5 and 50%+ improvement over DeepSeek/Llama isn't marginal — it's the difference between viable and unviable for cost-sensitive markets like India.
3. Sovereign AI is a feature, not a slogan. Every major market is developing AI sovereignty requirements: China with its Great Firewall, Europe with GDPR, India with its data localization laws. Artha is one of the first stacks built from day one for self-hosting on single nodes. That's not an afterthought — it's the product strategy.