What happened
MediaNama reports that Gnani.ai launched Artha at Upa-Rashtrapati Bhavan in New Delhi. The stack combines Evon v3.3, a 30-billion-parameter open-weight model trained in 11 Indian languages, with Plexus, a platform for developing and deploying AI agents. The report says model weights are available by request through Hugging Face under an Apache 2.0 licence.
MediaNama reports that Bengaluru-based voice AI company Gnani.ai introduced Artha at Upa-Rashtrapati Bhavan in New Delhi, with Vice President C.P. Radhakrishnan launching the stack. The product combines two components: Evon v3.3, described as a 30-billion-parameter multilingual language model trained in 11 Indian languages, and Plexus, an agentic platform intended to help organisations build and deploy agents for specific workflows. The report says Evon provides the core language capability while Plexus connects that capability to organisational tasks. These details come from MediaNama’s report and the statements it cites; they have not been independently confirmed here.
MediaNama reports that Gnani.ai is presenting Artha as a response to three adoption problems identified by the company: control over data, the cost of scaling AI and the difficulty of operating effectively across Indian languages and real-world scenarios. The report says Evon’s model weights are available on request through Hugging Face under an Apache 2.0 licence. That arrangement gives users a stated licensing framework, but access is not described as an unrestricted public download. The article also distinguishes between Evon and Plexus, meaning Artha is a combined model-and-platform offering rather than only a newly released language model.
The launch was accompanied by political messaging about technological self-reliance. According to MediaNama, Radhakrishnan said the combination of Evon and Plexus reflected the strength of India’s technology ecosystem and the ability of Indian engineers to develop advanced technologies. The report places Artha within India’s wider sovereign-AI push, which is motivated by concerns over dependence on US and Chinese infrastructure for sensitive data and by the expense of operating large foreign models. MediaNama also reports that Gnani.ai’s earlier 14-billion-parameter voice model received subsidised GPU compute through the IndiaAI Mission, but the source does not establish whether the same support applies to Artha or Evon v3.3.
Source details: medianama.com ↗
Why it matters
Artha is positioned as an India-based option for enterprises and government organisations concerned about foreign AI infrastructure, data control, scaling costs and support for Indian languages. Its practical significance remains unproven: MediaNama reports that prospective customers are still experimenting with the model and that Evon has not yet been deployed.
MediaNama’s account matters because it describes a concrete attempt to package an India-focused language model with an enterprise agent platform. A model trained across 11 Indian languages could be relevant to organisations whose customers, employees or records are not well served by English-first systems, while an agent platform could make the model usable in operational workflows rather than only in demonstrations. The report does not provide comparative language evaluations, accuracy measurements or evidence that Evon outperforms foreign alternatives, so those potential advantages remain claims or open questions.
The sovereignty argument also has practical limits. MediaNama reports that Artha is intended to reduce reliance on foreign AI infrastructure, but the source does not specify where every part of the system runs, what cloud or hardware dependencies remain, whether customer data stays in India, or how the company handles retention and access. An Indian company and an India-based product do not by themselves establish data residency, operational independence or security. Those issues will be important for government bodies, banks, insurers and other institutions handling sensitive information.
The licensing and access model deserves close attention. MediaNama notes that Evon’s weights are available only by request through Hugging Face, even though the reported licence is Apache 2.0. That can support commercial use while allowing the publisher to control distribution, but it is narrower than a fully public release and may limit independent auditing or reproducibility. The article’s broader point is that “open weight” should not automatically be treated as synonymous with fully open-source. The source does not state how requests are assessed, whether the weights are complete, or whether the training data, code and evaluation materials are available.
What to watch next
The decisive tests are independent evaluations, deployment in regulated workflows and clarity about data handling, security, pricing and access to the model weights. MediaNama reports early interest from five of 20 enterprises at a Pune meeting, but the article does not identify those organisations or document production results.
The first signal will be whether interest becomes verifiable production use. MediaNama reports that five of 20 enterprises at a recent Pune customer meeting expressed interest and began building on Evon, but also says that Evon had not yet been deployed. The report does not name the organisations, disclose contract terms or provide evidence from a live system. Future reporting should establish whether any bank, insurer, payments company or public institution has moved from experimentation to an operational deployment and under what human-review controls.
Regulated workflows will be a particularly important test. MediaNama reports that Gnani.ai co-founder Ganesh Gopalan said some voice-AI workflows are mostly autonomous, while underwriting and payments reconciliation require additional oversight before agents can fully implement them. That distinction indicates that the company is not claiming unrestricted autonomy in high-stakes settings. What remains unknown is how Plexus records agent decisions, handles errors, escalates uncertain cases, prevents unauthorised actions and supports audits. The source provides no independent safety assessment, security review, incident history or reliability data.
Independent technical evidence will determine how much weight to give the launch. Useful follow-up would include evaluations across the 11 reported Indian languages, comparisons with relevant open and proprietary models, costs at different usage levels, latency and hardware requirements, and documentation of the model’s training and testing. It will also be important to clarify whether Artha is generally available, how quickly weight requests are approved, and what support customers receive. MediaNama’s report establishes a launch and early commercial outreach, but it does not establish broad availability, production performance, customer impact or independent confirmation of the company’s claims.