An Indian-language assistant has to work where its users are. That could mean a connected office, a factory with restricted networks or a service counter where sending every request to the cloud is impractical.
Three models released under the IndiaAI Mission on November 3 bring different approaches to that problem: Soket Labs’ EKA-1 and ZenteiQ’s Dhīr 2B and Dhīr 5B. Their significance lies in the deployment choices they are designed to offer, while their developers’ claims still need testing against real work.
The Breakdown
EKA-1 is described as a 26.3-billion-parameter model trained on approximately 7.5 trillion tokens. Parameters are the numerical values a model learns during training; tokens are the chunks of text it processes. Neither number alone establishes the quality of its answers.
The announced language coverage spans all 22 scheduled Indian languages and English. Its stated focus includes mathematics, coding, reasoning and agentic applications, where an AI system carries out steps in a task.
Dhīr takes a smaller-model approach. ZenteiQ describes the two models as supporting English and ten Indian languages, with offline operation on local devices among their intended uses. Running a model locally can reduce dependence on an internet connection for each request. It does not by itself establish that an answer is accurate or that the surrounding application is secure.
What the deployment details reveal
ZenteiQ lists a compressed, four-bit Dhīr 2B model file of 1.21 GB. File size is not the same as the total memory an application needs: software, stored conversation and other components also matter.
The developer’s full technical report was still listed as forthcoming when checked. Its product page nevertheless identifies important limits. Dhīr is text-only; its instruction-following evaluation covers English only. It describes unsupported factual recall as a weak area and recommends authoritative reference material and human review for consequential uses.
Those disclosures matter for an Indian business considering a local assistant. Language coverage is a reason to test a model on regional-language work, not proof that it will handle every dialect, instruction or specialist term equally well.
The Big Picture
Consider a hypothetical factory assistant answering questions from an approved maintenance manual. A local model could be useful where connectivity is limited. But the evaluation would still need to establish whether it retrieves the right passage, follows the requested language and declines to invent an answer when the manual is silent.
That is a different purchasing decision from choosing the model with the largest parameter count. The relevant questions are what the application must do, where its data may go and what happens when its answer is wrong.
The Crunch
The new releases widen the set of Indian AI models worth evaluating. The next meaningful evidence will come from documented tests in the languages, devices and workflows where people actually use them.
Related category: Business & Brands
Leave your thoughts