Gnani.ai launches Artha, a sovereign AI stack for Indian enterprises and public institutions.
The AMW Read
Gnani.ai's first LLM release, framed as sovereign AI infrastructure and unveiled by India's vice president, updates the Indic-model roster without resolving an open debate.
Gnani.ai launches Artha, a sovereign AI stack for Indian enterprises and public institutions.
Voice AI startup Gnani.ai unveiled Artha, an end-to-end sovereign AI stack, at an event in New Delhi headlined by India's vice president, CP Radhakrishnan. Artha combines Evon v3.3, a 30-billion-parameter multilingual model trained on 2 trillion tokens across 11 Indian languages, with Plexus, Gnani.ai's agentic platform for building workflow-specific AI agents. Evon v3.3 is the company's first large language model release; its weights are available on request via Hugging Face under an Apache 2.0 license. Gnani.ai says the model uses roughly 40% fewer tokens than comparable models on Indian-language workloads, and plans to scale the family to 70-billion- and 100-billion-parameter versions while extending language coverage from 11 to 22.
The launch positions Gnani.ai as one of the few Indic-language model builders shipping an open-weight LLM rather than only wrapping third-party APIs, at a moment when sovereign control over data and infrastructure is becoming a distinct procurement requirement for Indian government and enterprise buyers. Pairing the model with an in-house agent layer also marks a shift from Gnani.ai's original voice-AI niche toward a broader stack play aimed at public-sector and enterprise deployments that favor domestically hosted alternatives.
Per the AI Market Watch index, Gnani.ai has raised $15.68M in total funding to date (coverage limited to companies tracked in our index) — modest capital for a company now fielding a 30B-parameter foundation model, its own agent framework, and a roadmap toward 100B parameters. It raised a $10M Series B in June alongside a voice-to-voice LLM launch, and Artha extends that trajectory into general-purpose language modeling within two months — a pace that will need either fresh capital or a lean training approach to sustain, and one investors and rival Indic-model builders should track closely.
