India has crossed a decisive technological threshold in artificial intelligence sovereignty. In a landmark announcement hosted in New Delhi, the Ministry of Electronics and Information Technology (MeitY), operating through the sovereign mandate of the IndiaAI Mission, officially unveiled three indigenous foundation AI models: EKA-1, Dhīr 5B, and Dhīr 2B. The releases represent India’s first coordinated, state-backed push to break structural dependencies on foreign AI application programming interfaces (APIs) controlled by US and Chinese hyperscalers, transitioning the domestic ecosystem from downstream consumers to primary foundation model architects.
At the center of the rollout is EKA-1, a 26.3-billion parameter sparse hybrid foundation model developed by New Delhi-based Soket Labs Technology and Research. Trained across 7.5 trillion tokens of multilingual data, EKA-1 natively supports all 22 scheduled Indian languages alongside English, delivering mathematical synthesis, complex code generation, and multi-step agentic workflows. Complementing this enterprise heavy-lifter are Dhīr 5B and Dhīr 2B, two compact, open-source small language models (SLMs) engineered by Bengaluru deep-tech startup ZenteiQ AiTech Innovations. Designed for 100% offline on-device execution, the Dhīr family brings native Indic reasoning directly to consumer smartphones and air-gapped enterprise infrastructure, eliminating cloud subscription tolls and preserving strict data privacy.
Sovereign Compute: Institutional Architecture of the IndiaAI Mission
The release of EKA-1 and the Dhīr lineup marks the first concrete software dividends of the Union Cabinet’s ₹10,372-crore IndiaAI Mission. For years, Indian software developers and domestic enterprises operated under an acute compute asymmetry. Building foundation models requires tens of thousands of high-bandwidth memory GPUs (such as NVIDIA H100 and B200 clusters), resources historically concentrated in Silicon Valley and Shenzhen. As geopolitical AI trade controls tighten and frontier models gatekeep proprietary weights, domestic digital infrastructure faced existential risks of technological vassalage.
To circumvent this bottleneck, the IndiaAI Innovation Centre established a dedicated Foundation Models pillar, offering selected deep-tech startups subsidized access to high-performance computing infrastructure, curated public sector datasets, and institutional grants. The formal release of the models in New Delhi was led by MeitY Secretary S. Krishnan, signaling state commitment to embedding sovereign AI across banking, citizen services, defense telemetry, and consumer hardware.
Under this state-subsidized compute allocation, Soket Labs and ZenteiQ AiTech Innovations were granted access to national compute fabrics and sovereign cloud partnerships, including Google Cloud TPU infrastructure and supercomputing clusters administered by the Centre for Development of Advanced Computing (C-DAC) in Pune. This institutional backing allowed the engineering teams to bypass the crushing capital expenditure barriers that have historically kept independent Indian AI research confined to fine-tuning foreign open weights.
EKA-1 Technical Architecture: 26.3B Parameters on Sparse Hybrid RALSNet
Unlike conventional dense transformer models that activate every neural parameter for every single token pass, EKA-1 is engineered around a sparse hybrid architecture powered by RALSNet. Spanning 52 layers, the model combines standard multi-head self-attention with recurrent linear attention mechanisms and sparsely routed Mixture-of-Experts (MoE) layers. This architectural synthesis yields three decisive operational advantages:
- Sub-Linear Memory Scaling: By interleaving recurrent linear attention layers with quadratic self-attention, EKA-1 slashes the key-value (KV) cache memory footprint during long-sequence generation, allowing longer document processing without memory exhaustion.
- Dynamic MoE Routing: The sparse expert layers route incoming tokens dynamically to specialized parameter clusters (such as formal mathematics, Python/Rust code synthesis, or Indic syntactic morphology), drastically reducing active FLOPs per forward pass while retaining the representational capacity of a 26.3B system.
- High-Throughput Inference: Lower active parameter requirements translate to superior time-to-first-token (TTFT) metrics and higher tokens-per-second generation on standard enterprise server GPUs.
EKA-1 was pre-trained from scratch on approximately 7.5 trillion tokens of highly curated text, code, and scientific literature. Crucially, the training corpus was balanced to provide native syntactic fluency across all 22 Eighth Schedule languages of the Indian Constitution (including Hindi, Bengali, Telugu, Marathi, Tamil, Gujarati, Kannada, Malayalam, Odia, Punjabi, and Assamese) alongside English. With an integrated 64,000-token (64K) context window, EKA-1 is purpose-built to parse voluminous Indian statutory filings, financial audit ledgers, and multi-file codebases in a single context ingestion.
Furthermore, Soket Labs designed EKA-1 to excel at agentic orchestration. The model natively interfaces with Soket’s open-source Loop agentic framework, executing autonomous API calls, multi-turn reasoning chains, and deterministic tool usage across sensitive domains like banking, financial services, insurance (BFSI), and enterprise cybersecurity telemetry.
Dhīr 2B and Dhīr 5B: Local Edge AI and Offline Smartphone Privacy
While EKA-1 tackles enterprise and sovereign server workloads, Bengaluru-based ZenteiQ AiTech Innovations addressed the opposite end of the computing spectrum: local, edge-native, zero-cloud inferencing. The Dhīr family consists of two distinct open-source models trained from the ground up on Google Cloud TPUs:
- Dhīr 2B (Dense SLM): A 2-billion parameter dense model engineered specifically for resource-constrained consumer hardware, including mid-range Android smartphones, tablets, and lightweight edge gateways. Quantized to INT4 and INT8 precisions, Dhīr 2B runs comfortably inside 1.5GB to 2GB of RAM, executing fully offline on modern mobile Neural Processing Units (NPUs) like MediaTek APUs and Qualcomm Snapdragon NPUs.
- Dhīr 5B (Sparse MoE SLM): A 5-billion parameter sparse Mixture-of-Experts model designed for local workstations, laptops, and secure enterprise servers. It provides enhanced multi-turn reasoning and domain classification while restricting active compute overhead to lightweight execution envelopes.
The philosophical foundation of the Dhīr project is absolute privacy and zero cloud dependence. Standard cloud-based AI assistants require users to upload confidential prompts, personal documents, and financial records to remote foreign data centers, incurring recurring operational costs and exposing metadata to international surveillance. In contrast, Dhīr models operate completely severed from the internet. Once the model weights are loaded into local system memory, all inference occurs strictly on-device.
For users running local language models on personal computers or workstation rigs, developers can deploy Dhīr weights seamlessly using open-source runtimes like llama.cpp and Ollama. If you encounter hardware driver conflicts where local models default to processor cores instead of dedicated graphics silicon, refer to our step-by-step explainer on how to fix Ollama running on CPU instead of GPU to unlock full hardware-accelerated inference speeds.
ZenteiQ has released both Dhīr 2B and Dhīr 5B under the permissive Apache 2.0 license, making model weights, tokenizers, and deployment scripts publicly downloadable through the government’s sovereign AIKosh portal and Hugging Face. Commercial enterprises, academic researchers, and independent app developers can freely inspect, modify, fine-tune, and embed the models into proprietary commercial applications without licensing fees or revenue-share clawbacks.
Architectural Breakdown: Sovereign Models Compared to Global Baselines
To understand where EKA-1 and Dhīr fit within the global AI landscape, consider their hardware requirements, parameter density, and linguistic capabilities against leading international benchmarks like Meta’s Llama 3.1 8B, Google’s Gemma 2 9B, and Mistral NeMo 12B:
| Model Name | Parameters | Architecture Type | Indic Language Support | Context Window | Target Deployment | Licensing Model |
|---|---|---|---|---|---|---|
| EKA-1 | 26.3 Billion | Sparse Hybrid RALSNet (52 Layers) | 22 Scheduled Languages + English | 64,000 Tokens (64K) | Enterprise Server & Cloud | Sovereign Open Weights (AIKosh) |
| Dhīr 5B | ~5 Billion | Sparse Mixture-of-Experts (MoE) | 10 Major Indic Languages + English | Optimized Edge Context | Workstations & Air-Gapped Servers | Apache 2.0 Open Source |
| Dhīr 2B | 2 Billion | Dense Small Language Model (SLM) | 10 Major Indic Languages + English | Edge Efficient Context | Smartphones & Local Edge NPUs | Apache 2.0 Open Source |
| Llama 3.1 8B | 8.0 Billion | Dense Decoder-Only Transformer | Limited Indic (English Dominant) | 128,000 Tokens (128K) | Cloud Server & High-VRAM PCs | Llama 3.1 Community License |
| Gemma 2 9B | 9.2 Billion | Dense Alternating Sliding Window | Secondary Indic Support | 8,192 Tokens (8K) | High-End Workstations | Gemma Terms of Use |
| Mistral NeMo | 12.2 Billion | Dense Decoder (Tekken Tokenizer) | Moderate Indic Support | 128,000 Tokens (128K) | Workstation & Cloud GPUs | Apache 2.0 Open Source |
Indic Token Economics: Eliminating the Unfair Token Tax
Perhaps the most revolutionary technical breakthrough delivered by EKA-1 and Dhīr lies in tokenization efficiency. For years, Indian enterprise adoption of generative AI has been crippled by the hidden “Indic Token Tax” imposed by Western foundational models.
Modern Large Language Models do not read whole words; they slice incoming text into sub-word tokens using algorithms like Byte-Pair Encoding (BPE). Western models like OpenAI’s GPT-4, Anthropic’s Claude, and Meta’s Llama are trained overwhelmingly on English corpora. In English, a standard sentence averages approximately 1.1 to 1.3 tokens per word. Common words like “technology” or “governance” are recognized as a single token.
However, when those same models process complex phonetic Indian scripts like Devanagari, Tamil, Telugu, Kannada, or Bengali, their tokenizers fail to recognize morphological roots. Instead, they fragment a single Indic word into 4 to 7 individual byte-level chunks. A simple 10-word Hindi sentence that should consume 13 tokens often registers as 60 tokens in GPT-4. This structural flaw creates severe computational distortions:
- Quadrupled Latency: Because autoregressive models generate text token-by-token, generating an Indic response requires 4x more sequential forward passes through the GPU, causing sluggish response times.
- Financial Penalty: Commercial API providers bill customers strictly per token consumed and generated. Indian businesses developing vernacular applications have routinely paid 300% to 500% higher API invoices than Western competitors asking identical questions in English.
- Context Window Erosion: A 128,000-token context window that comfortably holds a 100,000-word English document compresses to barely 20,000 words when ingesting Hindi or Tamil legal filings.
Both Soket Labs and ZenteiQ solved this imbalance by engineering proprietary multilingual tokenizers built specifically for Indian phonetic scripts. By expanding the native Indic vocabulary to include common ligatures, prefixes, and grammatical suffixes, EKA-1 and Dhīr achieve near 1:1 token-to-word parity (averaging 1.2 to 1.5 tokens per Indic word). This breakthrough slashes generation latency by up to 75%, equalizes API compute economics, and allows domestic enterprises to process extensive vernacular documents without context truncation.
On-Device Edge Deployment vs Cloud AI Subscription Traps
The strategic value of Dhīr 2B running locally on handheld silicon comes at a pivotal juncture in mobile computing. Smartphone OEMs and software platforms are aggressively steering consumers toward cloud-tethered artificial intelligence subscriptions. We have already observed major hardware manufacturers exploring paid recurring tiers for generative photo editing and smart summarization tools, as detailed in our analysis of Samsung testing paid Galaxy AI subscriptions once free promotional windows expire.
By releasing highly capable 2B and 5B parameter weights under an unencumbered Apache 2.0 license, India is establishing an open-source hedge against hardware monetization traps. A smartphone equipped with Dhīr 2B can perform real-time local voice transcription, offline document summarization, secure on-device SMS spam classification, and vernacular language translation without pinging an external server, without transmitting biometric or financial data, and without demanding monthly subscription fees.
Furthermore, in critical defense applications, tactical border communications, and air-gapped banking intranets, continuous internet access is either impossible or strictly prohibited due to national security regulations. Dhīr models can be compiled directly into secure firmware chips, localized edge nodes, and field radios, ensuring operational continuity even during wide-area network blackouts or satellite jamming events.
Strategic Outlook: AIKosh, Compute Scaling, and Long-Term Viability
The arrival of EKA-1, Dhīr 5B, and Dhīr 2B proves that India’s deep-tech engineering talent can architect world-class foundation models when compute bottlenecks are alleviated by institutional policy. However, maintaining sovereign technological independence requires sustained long-term execution rather than celebratory launch milestones.
As search engines and global information retrieval increasingly shift toward generative reasoning platforms, keeping pace with international foundation models will be critical. With recent global web analytics revealing that ChatGPT has climbed to the top 5 most visited websites globally alongside Google, the battleground for digital mindshare is decisively transitioning to sovereign natural language understanding.
For India to scale this initial triumph into lasting global competitiveness, three immediate challenges must be addressed:
- Accelerating Sovereign Hardware Procurement: While the subsidized compute access granted to Soket Labs and ZenteiQ proved catalytic, MeitY must expedite the broader procurement and deployment of its planned 10,000+ GPU sovereign cloud cluster to prevent researchers from queuing for compute allocations.
- Curating High-Quality Vernacular Pre-Training Corpora: High-parameter models are only as robust as their pre-training data. The IndiaAI Datasets Platform (AIKosh) must aggressively digitize, sanitize, and publish open-access public archives across rural dialects, regional legal case law, and technical curricula.
- Domestic Enterprise Adoption Mandates: Indian public sector units (PSUs), domestic banking consortiums, and state education departments must actively integrate EKA-1 and Dhīr into their production pipelines rather than continuing to funnel taxpayer capital into foreign proprietary APIs.
By releasing open weights, prioritizing native Indic tokenization, and targeting both high-end server clusters and local offline smartphones, India has demonstrated a pragmatically differentiated AI strategy. EKA-1 and Dhīr are not mere academic proofs of concept: they are the operational cornerstones of India’s sovereign digital future.