Gnani opens a 30B model for India's 11 languages

Gnani launched Artha with Evon v3.3, a 30B-parameter open-weight model with roughly 3.5B active parameters trained across 11 Indian languages, and says it can run on one inference node. Gnani’s launch page lists Apache 2.0 weights, self-hosted deployment, and about 20% fewer tokens per Indian word than the GPT-5 family tokenizer; the model is also available by request on Hugging Face. Those efficiency comparisons are company-reported, so Indian builders should benchmark the sovereign stack against the competition among frontier labs for sovereign AI before treating language coverage or residency as a reason to switch.