Two government-backed startups release sovereign AI models that can run in airgapped environment

An air-gapped AI system operates on a computer or network completely cut off from the internet and outside networks. A sovereign AI model, in this article’s context, is one released by Indian startups for secure, local and offline use. Together, these ideas address situations where organisations cannot send information to remote AI services.

For example, Soket AI’s EKA-1 and ZenteiQ’s Dhīr models can perform inference locally. Inference is the step where an AI model produces an answer. Unlike a typical online chatbot, the local system does not need to contact a distant server. ZenteiQ said no data will go outside the air-gapped setup.

The models were backed by the government’s India AI Mission and presented for use in coding, maths, science and research. Defence and cybersecurity were named as possible users. Their offline design could also support critical sectors that require tighter control over data and network access.

Soket AI launched EKA-1, a 26.3 billion parameter model designed mainly for mathematics and coding. Its training covered all 22 scheduled Indian languages and English. The model is intended to bring substantial maths and programming capability into a system that can operate offline.

ZenteiQ released two open-source models: Dhīr 2B and Dhīr 5B. They are multilingual assistants supporting English and 10 Indian languages. ZenteiQ designed them for secure, on-premise and offline deployment, rather than requiring an internet-connected service.

The startups presented the models as tools for sensitive or specialised work. IT Ministry secretary S Krishnan described Soket’s model as focused on maths and coding, while ZenteiQ’s models were aimed at science and research. Defence and cybersecurity were also identified as possible uses for smaller offline models.

EKA-1 contains 26.3 billion parameters, the numerical settings an AI model learns during training. Soket AI trained it on about 7.5 trillion tokens. The training material covered all 22 scheduled Indian languages plus English, giving the model coverage across 23 languages.

ZenteiQ’s two models are smaller: Dhīr 2B has 2 billion parameters, and Dhīr 5B has 5 billion. Together, they are described as multilingual assistants supporting English and 10 Indian languages. The article does not provide their training-data volume.

These figures show different design priorities. EKA-1 is larger and focused on maths and coding, with extensive multilingual training. Dhīr models use smaller sizes and are designed for local, offline deployment. ZenteiQ said its models can run on a mobile phone with at least 6 GB of memory, widening their possible reach.

Inference is the moment an AI model processes a request and generates an answer. In a service such as Axiom, ZenteiQ explained that this step happens on a remote server. The user’s request therefore reaches a server outside the local device or organisation.

In an air-gapped system, inference happens on local hardware. The model is installed within the isolated environment and does not need an internet connection to respond. ZenteiQ’s Sashikumar Ganesan said that no data will go outside. He also said the model can run on a mobile phone with at least 6 GB of memory.

This difference matters for organisations handling sensitive information. Local inference keeps the processing inside the protected environment, while a remote chatbot depends on communication with an external service. The article names defence and cybersecurity as possible users of this offline approach.

When an AI model runs locally without internet access, the information used for a request remains within the device or isolated network. The system does not need to transmit that request to a remote service for processing. This is the central privacy and security benefit described in the article.

ZenteiQ’s Sashikumar Ganesan explained that inference occurs locally in an air-gapped setup. He said the model needs no network connection and that no data will go outside. This contrasts with services such as Axiom, where inference happens on a remote server. ZenteiQ said its model can even run on a phone with at least 6 GB of memory.

Keeping processing local can help organisations work with sensitive information while limiting external exposure. The article specifically mentions defence and cybersecurity as possible users. It does not claim that local operation removes every security risk, but it does prevent the model from requiring an outside network connection.

Offline AI matters to sensitive sectors because it can operate without connecting their systems to the internet or outside networks. The article identifies defence and cybersecurity as possible users. It also describes the models as suitable for secure, on-premise use in critical sectors.

The key mechanism is local inference. Instead of sending a request to a remote server, the model produces its answer inside the air-gapped environment. ZenteiQ said no data will go outside and that its model needs no network connection. This arrangement is designed for organisations that need tighter control over where information is processed.

Research institutions and other critical-sector users could apply the models to science, research, maths or coding while keeping work local. The article does not list specific research institutions or detailed deployment plans. It does show that smaller models can support specialised tasks and may run on devices with at least 6 GB of memory.

GPUs, or graphics processing units, are processors that can perform many calculations in parallel, making them useful for AI training. TPUs, or Tensor Processing Units, are specialised processors designed for tensor calculations used by many AI systems. These definitions reflect established computing knowledge; the article mainly discusses their availability for training.

ZenteiQ’s Sashikumar Ganesan said the company trained its models on TPUs because GPUs were scarce. He said TPUs are not inferior to GPUs. Access to suitable hardware determines whether a team can begin training promptly and handle the large calculations involved in building models.

Ganesan said obtaining GPUs would have required waiting another three to four months. Using available TPUs therefore helped ZenteiQ avoid that delay. The example shows that AI development depends not only on model design and training data, but also on access to specialised computing resources and the practical limits of hardware supply.