Gnani AI introduced Artha, a sovereign AI stack for Indian institutions that need models that read local languages, stay inside their own systems, and still clear daily volume. The company presents the launch as a Vice-President of India event. At the center is Gnani Evon v3.3, a 30 billion parameter language and reasoning model with 3.5 billion active parameters, trained natively across 11 Indian languages.

Gnani says Evon v3.3 weights are open under Apache 2.0 and available by request on Hugging Face. The model is meant to run on a single node. The company rebuilt the tokenizer for Indian scripts and says that cut about 20 percent of the tokens per Indian word versus the GPT-5 family tokenizer. Fewer tokens, in its telling, are how Indic text stops being a tax on context length and cost.

A stack, not only a checkpoint

Artha packages four models with an agent platform called Plexus. Beside Evon v3.3, Gnani lists Evon v2.0 for enterprise reasoning and tool use, Prisma v2.5 for speech recognition on real telephone audio, and Timbre v2.5 for streaming speech synthesis. Plexus turns those models into identity-bearing agents that can run in sequence or in parallel, call existing systems, and sit under human or automatic review.

On the MILU Indian-language benchmark, Gnani says Evon v3.3 beats a 105 billion parameter Indic model on 10 of 11 languages and a similarly sized 30 billion parameter model on all 11, while matching a hosted global frontier model of similar size. It also says every model it benchmarked against Evon v3.3 costs at least 2.2 times more per point of Indian-language accuracy. Those comparisons are the company’s own.

The pitch to banks, insurers, and ministries is residency. Gnani says the stack can live in a customer’s data center or VPC so files never leave the network, a setup it maps to India’s DPDP law and to RBI and IRDAI expectations. Sample jobs on the Artha page run from loan-file underwriting and KYC cleanup to grievance routing, subsidy-failure calls, and disaster triage in the citizen’s own language.

Decoded Take

India has no shortage of “sovereign model” announcements. Artha is more interesting as a packaging argument: open weights, a cheaper Indic tokenizer, and an agent layer that is supposed to run next to core-banking and welfare systems rather than in a public API. The 30 billion parameter claim only matters if a lender can actually host Evon v3.3, keep the file on-prem, and beat the translation-plus-GPT bill on a real queue of Marathi applications. Watch whether the Hugging Face request path turns into unrestricted weights, and whether a named bank or state department says it put Plexus on production traffic. Until then, this is a well-aimed stack with company-run scores.