Yes. Dhīr AI models are designed to generate responses locally, without a continuous internet connection, once installed on compatible hardware. That could put a document assistant within reach of an office or classroom where connectivity repeatedly fails.
India’s Ministry of Electronics and Information Technology announced Dhīr 2B and Dhīr 5B, developed by ZenteiQ AiTech Innovations, on November 3, 2026. The IndiaAI Mission release also introduced Soket’s EKA-1. These are separate model families with different specifications.
Key Highlights
- Dhīr supports English and 10 Indian languages.
- Local deployment allows use without continuous cloud access.
- Two model sizes offer different computing requirements.
- Institutional applications still need testing against their intended tasks.
What India Released: Dhīr 2B and Dhīr 5B
The model names describe their approximate scale. ZenteiQ’s research abstract identifies Dhīr 2B as a dense transformer and Dhīr 5B as a mixture-of-experts model, which activates only part of its parameters for each token.
| Specification | Dhīr 2B | Dhīr 5B MoE |
| Total Parameters | 2.03 billion | 5.09 billion |
| Active Parameters Per Token | Dense architecture | 1.79 billion |
| Published 4-Bit File Size | 1.21 GB | 2.95 GB |
Both were trained from scratch. The research abstract reports more than six trillion pretraining tokens per model and training on Google Cloud TPUs in India.
A smaller download does not establish whether an institution’s existing computers will deliver acceptable response times. The software running the model and the workload also need assessment.
How Offline AI Could Help Schools, Hospitals and Government Offices
The government announcement identifies document-grounded automation and knowledge retrieval among Dhīr’s intended uses. A possible deployment would connect the model to approved local documents so answers draw on those materials.
For the settings in this headline, potential applications include:
- Schools: Answering questions from a locally stored collection of approved teaching material during connectivity outages.
- Hospitals: Helping staff locate administrative instructions or internal procedures, with source checking before use.
- Government offices: Retrieving information from departmental manuals or explaining a service process using authorised local documents.
These are illustrative possibilities, not confirmed installations. The launch announcement does not establish that Dhīr has been deployed across these institutions.
The distinction is especially important in hospitals. Finding an administrative document is a narrower task than recommending treatment. A hospital pilot would need to define that boundary before staff rely on its output.
What Institutions Need Before Switching Off the Internet
ZenteiQ’s published hardware tests show why device selection needs attention. Dhīr 2B generated 18.6 tokens per second on Intel Iris Xe integrated graphics, compared with 2.3 on a Raspberry Pi 4 CPU, using 4-bit weights and llama.cpp.
Those are developer-reported measurements for a single stream. They do not establish performance when several employees submit requests simultaneously.
A pilot therefore needs to answer three questions:
- Does the actual device have sufficient memory beyond the model file itself?
- Can staff identify the document supporting an answer and detect an incorrect response?
- Who maintains the local files, permissions, and model updates?
Offline operation also does not automatically supply current information. A locally stored circular remains outdated after a replacement appears elsewhere until someone updates the collection. For government services, that could change which procedure an assistant describes.
Dhīr AI Models: Language Support, Access and Remaining Limits
The supported languages are Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi and Odia, alongside English. That list should not be confused with EKA-1’s coverage of all 22 scheduled Indian languages.
ZenteiQ links model downloads through AIKosh and lists an Apache 2.0 licence. Availability of model weights does not remove the cost of equipment or deployment work.
Its documentation also flags weak factual recall without supporting sources and recommends human review for sensitive uses. Schools and offices should judge a pilot against their own documents rather than treating launch claims as proof of accuracy.
Frequently Asked Questions
Do Dhīr models accept voice commands?
The published models are text-only. ZenteiQ lists no audio, speech, or image input or output, so voice interaction would require an additional system around the model.
Does Indian-language support guarantee equal performance?
No. Language coverage does not establish equal accuracy across tasks. ZenteiQ says instruction following was measured in English only, leaving institutions to assess their chosen languages.
Are these ready-made mobile apps?
The announcement concerns foundation models. Installing weights requires compatible software and hardware; it does not automatically provide an institution with a finished application or support service.
How could a school test Dhīr before wider use?
A school could start with one approved subject collection and teacher-written questions. Comparing responses with source material would expose gaps before extending the pilot to other subjects.
Is the full technical report available?
The publications page currently provides a research abstract and says the technical report is coming soon. Readers should distinguish that abstract from a complete published report.
Explore More AI Stories Shaping India’s Future!
1. Are Chinese AI models facing distillation allegations?
Find out more about the reported DeepSeek and Moonshot AI model distillation concerns.
2. How will India regulate AI risks?
Explore India’s proposed AI safety measures and model risk checks for 2026.
3. Are AI skills becoming more valuable than degrees?
Discover why AI skills are gaining attention in India’s changing employment market.
4. Could AI create a new digital divide?
See how differences in AI knowledge and responsible usage could influence future opportunities.
5. Why is India investing ₹730 crore in quantum AI?
Uncover the reported ₹730 crore quantum AI university initiative and its potential significance.
