On November 3, 2026, the IndiaAI Mission announced indigenous foundation models EKA-1 from Soket and Dhīr 2B and Dhīr 5B from ZenteiQ. The government says EKA-1 has 26.3 billion parameters and was trained on about 7.5 trillion tokens spanning all 22 scheduled Indian languages plus English. Dhīr 2B and 5B are Apache 2.0-licensed models supporting English and 10 Indian languages, with ZenteiQ positioning them for on-premise and fully offline deployment.
Key takeaways
- The IndiaAI Mission announced EKA-1, Dhīr 2B and Dhīr 5B on November 3, 2026.
- EKA-1, developed by Soket, is described by PIB as a 26.3-billion-parameter sparse hybrid language model trained on about 7.5 trillion tokens.
- PIB says EKA-1 spans all 22 scheduled Indian languages plus English and is designed for math, code, reasoning and agentic use.
- Dhīr 2B has 2.03 billion parameters, while Dhīr 5B MoE has 5.09 billion total parameters with about 1.79 billion active per token.
- ZenteiQ says both Dhīr models were pretrained from scratch on more than 6 trillion tokens per model.
- Dhīr supports English plus 10 Indian languages and is released under the Apache 2.0 licence.
- ZenteiQ says Dhīr can run fully offline on laptops and edge hardware; exact performance depends heavily on device and quantisation.
- The government announcement does not establish that EKA-1 or Dhīr beats leading global frontier models, so parameter counts should not be confused with proven superiority.
- As of November 10, Soket’s own EKA-1 access page still describes access as early access via a waitlist, so “released” should not be interpreted as unrestricted public download availability.
What did IndiaAI announce on November 3, 2026?
On November 3, 2026, the IndiaAI Mission under the Ministry of Electronics and Information Technology announced three indigenous foundation-model releases: EKA-1, developed by Soket Labs Technology and Research Private Limited, and Dhīr 2B and Dhīr 5B, developed by ZenteiQ AiTech Innovations. The announcement was made in New Delhi under the IndiaAI Innovation Centre pillar. The official Press Information Bureau release says the companies were among the teams selected by IndiaAI to build indigenous foundation AI capabilities and large language models.
PIB also says the projects received access to subsidised national AI compute infrastructure and support for research expenses under the mission. The two model families target different parts of the market. EKA-1 is the larger model in the announcement, with 26.3 billion parameters and an emphasis on multilingual reasoning, math, code and agentic use.
Dhīr 2B and 5B are substantially smaller models designed around local deployment, lower hardware requirements and multilingual operation in environments where sending data to a remote cloud service may be undesirable or impossible. IndiaAI COO Sudeep Shrivastava described the releases as covering both ends of that spectrum: lightweight models optimised for everyday devices and larger systems intended for more demanding enterprise applications. That framing is useful because these models should not be treated as interchangeable simply because they were announced at the same event.
What is EKA-1?
EKA-1 is a foundation model developed by Soket Labs with support from the IndiaAI Mission. According to PIB’s November 3 release, EKA-1 has 26.3 billion parameters; approximately 7.5 trillion training tokens; coverage across all 22 scheduled Indian languages; English support; and a stated focus on math, code, reasoning and agentic use. In practice, that makes EKA-1 one of the headline models of the November 3 announcement.
PIB describes EKA-1 as an open, sparse hybrid language model. The word “sparse” is important. In sparse architectures, not every parameter necessarily participates equally in processing every token. That means parameter count alone does not translate directly into inference cost or model quality. It also means a simple comparison such as “26.3B is bigger than 5B, therefore it is five times better” would be meaningless. Model architecture, training data, post-training, evaluation, inference design and task requirements all matter.
Is EKA-1 publicly downloadable?
The November 3 government announcement uses the language of a model release, but Soket’s own current access information adds an important qualification. As of November 10, Soket’s EKA-1 early-access page asks developers to join a waitlist, after which Soket reviews the request and grants access to approved users. Soket’s broader Project EKA page describes the programme as its flagship research effort in reasoning and multilingual AI.
So the safest description is: EKA-1 has been officially announced/released under the IndiaAI Mission, but unrestricted public access should not be assumed. Soket currently directs prospective users through an early-access process. That distinction matters for developers searching for a model download, API endpoint or public weights.
How was EKA-1 trained?
The most specific public number in the government announcement is the training corpus size. PIB says EKA-1 was trained on approximately 7.5 trillion tokens. It also says the model spans all 22 scheduled Indian languages plus English. Soket’s wider Project EKA programme describes a larger research effort involving math, code, reasoning, multilingual AI, Indian and Global South languages and sovereign compute supplied through the IndiaAI Mission.
However, those broader Project EKA figures should not automatically be treated as specifications for the specific 26.3B EKA-1 release. For example, Soket’s project site discusses a much larger ongoing programme with architectures beyond 100 billion parameters and substantially more curated training data. Those numbers refer to the Project EKA programme, not necessarily to EKA-1 itself. For EKA-1, the verified release-specific figures remain 26.3B parameters and approximately 7.5T tokens.
Which languages does EKA-1 support?
PIB says EKA-1 covers all 22 languages listed in the Eighth Schedule of the Indian Constitution, plus English. That includes India’s constitutionally scheduled languages rather than only the few high-resource Indian languages most commonly supported by commercial AI products. This gives EKA-1 an explicitly India-focused multilingual positioning. The significance is not simply the number 22. Indian-language AI faces difficult technical problems involving different scripts, code-mixing with English, romanised Indian-language text, Limited high-quality training data for some languages and large differences in online data availability between languages.
A model trained across all scheduled languages is therefore potentially useful for building applications intended to work beyond English and Hindi. However, coverage does not automatically mean equal quality. Neither PIB nor Soket’s public release information establishes that EKA-1 performs equally well in all 22 Indian languages. There is also no official language-by-language benchmark in the PIB announcement proving superiority over other models. The safe conclusion is that EKA-1 was developed with all 22 scheduled Indian languages in its training and positioning — not that every language necessarily receives identical performance.
What can EKA-1 do?
PIB names four broad capability areas mathematics, code, reasoning and agentic use. That positions EKA-1 as more than a multilingual chatbot. Soket’s own research programme similarly focuses heavily on mathematical reasoning, software generation and long-horizon agentic workflows. The company describes Project EKA as an effort to develop models that can work through math, code and logic rather than simply generate plausible text. Still, public claims need to be kept separate from measured performance. As of November 10, the official IndiaAI announcement does not provide enough independent benchmark data to conclude that EKA-1 outperforms major international frontier models, leads every Indian-language benchmark, has a particular reasoning score, beats specific commercial systems, has a fixed public API price, runs efficiently on a particular consumer GPU and offers a particular context-window length. Those details should come from Soket’s technical documentation or independently reproducible evaluations rather than being inferred from parameter count.
What are Dhīr 2B and Dhīr 5B?
Dhīr is a family of compact foundation models developed by ZenteiQ. Both are designed for deployment on modest hardware rather than only in large data centres. ZenteiQ publishes the architecture details and licence for both on its model page. The two released variants are:
| Model | Architecture | Total parameters | Active parameters |
|---|---|---|---|
| Dhīr 2B | Dense transformer | 2.03B | 2.03B |
| Dhīr 5B MoE | Mixture of Experts | 5.09B | ~1.79B per token |
These specifications come from ZenteiQ’s official Dhīr model page. The 5B model uses a Mixture-of-Experts, or MoE, architecture. ZenteiQ says it contains 12 experts and routes each token through two of them. That allows the model to hold more total parameters while using only a subset for each token.
Both models have a 16,384-token context window after post-training, according to ZenteiQ. The company says each model saw more than 6 trillion tokens during training. The context window is the amount of text a model can consider at once, so 16K tokens suits short documents and conversations rather than very long files.
Were Dhīr models trained from scratch?
Yes, according to ZenteiQ. Its technical training overview says both Dhīr models were pretrained from random initialisation rather than being fine-tuned versions of another developer’s foundation model. ZenteiQ says the training was performed on Google Cloud TPU v6e hardware in Mumbai, with data and checkpoints maintained in the same region. ZenteiQ’s reinforcement-learning write-up says it subsequently used RL to improve instruction following, tool use, multilingual responses and safety behaviour. Those are developer-published technical claims, so they are more detailed than the government’s announcement but should still be distinguished from independent third-party verification. The company also says its training stack included:
- More than 6 trillion pretraining tokens per model
- English web text
- Code
- Mathematics and reasoning data
- Multilingual material
- 10 Indian languages
- Long-context training to 16K tokens
- Supervised fine-tuning
- Reinforcement learning
Which languages does Dhīr support?
ZenteiQ says Dhīr supports English plus 10 Indian languages Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi and Odia. Its documentation says these languages were included in pretraining rather than added only during fine-tuning. The company also says its corpus includes both native-script and romanised forms of Indian-language text. That is relevant in India because users frequently mix English and Indian languages or type Indian languages using Latin characters. Dhīr therefore has narrower language coverage than EKA-1’s stated 22 scheduled Indian languages, but the two products are targeting different model sizes and deployment scenarios.
Are Dhīr 2B and 5B open source?
Yes. ZenteiQ’s current Dhīr model page lists the models under the Apache License 2.0. The page also links to AIKosh for model access and says the models are available as instruction-tuned checkpoints.
That is more specific than simply calling the models “open.” Apache 2.0 is a widely used permissive software licence that generally allows use, modification and redistribution subject to its licence conditions. Developers should still read the actual licence and model documentation before deployment, particularly for regulated or commercial applications.
Can Dhīr really run offline?
Yes — offline deployment is one of ZenteiQ’s central design goals for the models. The company’s official Dhīr page states that the models require no network connection for local inference once deployed. ZenteiQ positions them for laptops, embedded systems, edge devices, industrial environments, network-denied settings, public-service kiosks and Private local inference. That makes Dhīr different from cloud-only AI services where every prompt must be sent to a remote data centre. ZenteiQ lists the following sizes:
| Model | 4-bit file | 16-bit weights |
|---|---|---|
| Dhīr 2B | 1.21 GB | 4.06 GB |
| Dhīr 5B MoE | 2.95 GB | 10.19 GB |
Those figures make local deployment more plausible than it would be for a much larger frontier model. However, file size does not tell you whether the experience will be fast enough on every computer. Inference performance depends on CPU/GPU architecture, available RAM, memory bandwidth, quantisation, runtime, context length and prompt size. So “offline-capable” should not be translated into “runs well on every phone.”
What hardware can Dhīr run on?
ZenteiQ has published specific test results for several devices. Its model page reports 4-bit llama.cpp inference on hardware including intel Iris Xe integrated graphics, NVIDIA Jetson Orin NX and raspberry Pi 4. The company reports single-stream decode rates in its own testing, but those figures are hardware-, software- and configuration-specific. They should not be interpreted as guaranteed performance for every laptop or edge computer. More importantly, ZenteiQ’s public documentation does not establish that every smartphone can practically run Dhīr.
The company talks primarily about laptops, embedded boards and edge hardware. There is also a ZenteiQ mobile application called ZiqeX listed on the company’s website, but that should not be confused with a claim that the raw Dhīr 2B or 5B model weights run locally on every consumer phone. Offline-capable deployment and mainstream smartphone availability are different questions.
What is the difference between EKA-1 and Dhīr?
The simplest comparison is scale versus deployment flexibility. EKA-1 is a large government-announced foundation model, while Dhīr is a family of small models built to run locally. The table below sets out the main differences.
| Feature | EKA-1 | Dhīr 2B | Dhīr 5B MoE |
|---|---|---|---|
| Developer | Soket | ZenteiQ | ZenteiQ |
| Parameters | 26.3B | 2.03B | 5.09B total |
| Language positioning | 22 scheduled Indian languages + English | 10 Indian languages + English | 10 Indian languages + English |
| Training tokens | ~7.5T | 6T+ | 6T+ |
| Architecture | Sparse hybrid | Dense | Mixture of Experts |
| Core positioning | Math, code, reasoning, agents, multilingual AI | Compact local assistant | Compact MoE local assistant |
| Offline positioning | Not the central claim in PIB release | Yes | Yes |
| Public licence | Not clearly established for EKA-1 public weights | Apache 2.0 | Apache 2.0 |
| Current access | Soket early-access process | Released | Released |
EKA-1 is much larger and is positioned as a broader reasoning and multilingual foundation model. Dhīr is designed around being small enough to deploy locally. Neither approach is automatically “better.” A bank building a large reasoning system in a controlled data centre has different requirements from a factory that wants an assistant running without internet connectivity at the edge.
Who should care about EKA-1 and Dhīr?
The models are relevant to different groups for different reasons. Developers, enterprises and researchers will each look at different parts of the announcement. We take each group in turn below.
Developers
Developers building multilingual applications may care about both model families. EKA-1’s attraction is its claimed coverage of all 22 scheduled Indian languages plus English, combined with math, code and reasoning capabilities. Dhīr is potentially more immediately accessible to developers because ZenteiQ publishes concrete model specifications, an Apache 2.0 licence and an AIKosh download path. Its smaller model sizes also make experimentation possible without the infrastructure associated with very large models.
Enterprises
The strongest enterprise angle for Dhīr is local control. ZenteiQ explicitly offers deployment through cloud, Private cloud, on-premises infrastructure and fully offline edge environments. That can matter in organisations that cannot send sensitive information to an external cloud service. ZenteiQ specifically highlights industrial sites, strategic or network-denied settings and privacy-sensitive environments. Soket similarly positions its broader platform for enterprise and sovereign infrastructure, although current EKA-1 access remains more controlled.
Students and researchers
Researchers and students interested in Indian-language AI may find the releases important because both programmes were built around multilingual Indian data rather than treating it as an afterthought. Dhīr is particularly relevant for reproducibility because ZenteiQ publishes model architecture details, training information and an open licence. IndiaAI also says models can be made available through AIKosh, its national platform for AI datasets, models, toolkits and research resources. AIKosh describes itself as a national hub intended to improve access to datasets and models for innovation and research.
What do the Dhīr benchmarks actually show?
Unlike the brief government announcement, ZenteiQ has published evaluation results for Dhīr. Its model page reports evaluations across Indic-language, agentic, code/maths and English categories. ZenteiQ says the models were compared with other open models in the 1B-to-7B range under a common evaluation protocol. The company reports that:
- Dhīr 2B performs particularly strongly on its selected Indic benchmark suite.
- Dhīr 5B MoE performs better than Dhīr 2B in its agentic, code/maths and English averages.
- Dhīr 2B leads the 5B model on its aggregate Indian-language evaluation.
Those are developer-published benchmark results, not independent proof that Dhīr is the best model in its class. ZenteiQ itself also publishes limitations. It says factual recall without supporting sources is a weak area and recommends retrieval from authoritative sources plus human review for fields such as law, finance, medicine, agriculture and history. It also says larger models remain stronger in mathematics and code and explicitly states that the current Dhīr releases do not have a dedicated reasoning mode. That caution is useful because benchmark claims should be read alongside documented limitations.
What does the IndiaAI announcement not prove?
This is particularly important with AI releases, where technical announcements can quickly turn into exaggerated headlines. The November 3 announcement does not establish that EKA-1 is India’s objectively “best” AI model; EKA-1 beats GPT, Gemini, Claude or other international frontier models; dhīr outperforms every other compact model; every supported Indian language performs equally well; EKA-1 is freely downloadable by anyone today; dhīr runs effectively on every consumer smartphone; a larger parameter count automatically means a smarter model; and either model is suitable for unsupervised use in high-stakes sectors. The verified story is already substantial without those claims. IndiaAI has backed multiple home-grown foundation-model programmes with national compute infrastructure. Soket has produced a 26.3B-parameter multilingual model spanning 22 scheduled Indian languages plus English.
ZenteiQ has released 2B and 5B compact models with an open licence, published architecture details and an explicit emphasis on offline deployment. Those are concrete milestones. They do not need unsupported claims of global superiority to be significant.
Why does IndiaAI care about indigenous foundation models?
The broader goal of the IndiaAI Mission is to build domestic AI infrastructure and reduce dependence on external systems. PIB says EKA-1 and Dhīr were developed with access to subsidised national compute infrastructure provided through the mission. For India, foundation-model development has several strategic dimensions Indian-language coverage, domestic AI research capability, data sovereignty, local deployment, enterprise use, government applications and reduced dependence on foreign cloud AI platforms.
That does not mean an Indian model is automatically preferable to a global model for every task. The strategic value is that organisations gain more options. A model built and deployed domestically can be useful when language coverage, local infrastructure, offline inference or data-control requirements matter more than simply accessing the largest available cloud model.
FAQ
What is EKA-1?
EKA-1 is a foundation model developed by Soket under the IndiaAI Mission. PIB says it has 26.3 billion parameters, was trained on approximately 7.5 trillion tokens and covers all 22 scheduled Indian languages plus English.
When was EKA-1 announced?
IndiaAI officially announced EKA-1 alongside Dhīr 2B and Dhīr 5B on November 3, 2026. The announcement came through the IndiaAI Mission and a PIB release.
Is EKA-1 open source?
PIB describes EKA-1 as an open, sparse hybrid language model, but Soket’s current public page still directs users to an early-access waitlist. An unrestricted public-weights release and its exact licensing terms should therefore not be assumed from the government wording alone.
Can I download EKA-1 today?
Not freely. Soket currently directs prospective users through an early-access application. The company says applicants join a waitlist and receive access after review.
What is Dhīr?
Dhīr is ZenteiQ’s family of compact multilingual foundation models. The released versions are Dhīr 2B, a 2.03B-parameter dense transformer, and Dhīr 5B MoE, which has 5.09B total parameters with about 1.79B active for each token.
Are Dhīr models open source?
Yes. ZenteiQ currently lists Apache 2.0 as the licence for Dhīr 2B and Dhīr 5B MoE. That is a permissive open-source licence, but check ZenteiQ’s model page for the current terms.
How many languages does EKA-1 support?
PIB says EKA-1 covers all 22 scheduled Indian languages plus English. The language coverage is one of the main points of the government announcement.
How many languages does Dhīr support?
Dhīr supports English and 10 Indian languages: Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi and Odia. That is narrower than EKA-1’s coverage of all 22 scheduled languages.
Can Dhīr run offline?
Yes. ZenteiQ specifically designed Dhīr for fully offline inference. It says the models can run on laptops and edge hardware without a network connection once deployed.
Can Dhīr run on a normal laptop?
ZenteiQ says yes for supported configurations. Its 4-bit Dhīr 2B file is about 1.21 GB, while the 4-bit Dhīr 5B MoE file is about 2.95 GB, and actual speed depends on the laptop’s hardware and inference software.
Can Dhīr run on a smartphone?
ZenteiQ’s published material focuses on laptops, embedded boards and edge devices. Offline capability should not be interpreted as confirmation that the raw models run effectively on every consumer phone.
What is the context window of Dhīr?
ZenteiQ lists a 16,384-token context window for both released Dhīr models after training and post-training. In practice that suits short documents and conversations.
Were Dhīr models trained from scratch?
Yes, according to ZenteiQ. The company says both models were pretrained from random initialisation rather than being derived from another developer’s existing checkpoint.
Where were Dhīr models trained?
ZenteiQ says the models were trained on Google Cloud TPU v6e infrastructure in Mumbai. Data and checkpoints were kept in the same region.
Does EKA-1 beat GPT, Gemini or other frontier models?
The official IndiaAI announcement does not provide evidence supporting that conclusion. EKA-1’s parameter count and multilingual coverage are confirmed, but direct superiority over particular global frontier models should not be claimed without comparable independent benchmarks.
Is EKA-1 better than Dhīr?
They target different requirements. EKA-1 is a much larger multilingual model positioned around reasoning, code, math and agentic workloads, while Dhīr is deliberately smaller and designed for local, offline and edge deployment. Which is more suitable depends on the use case.
