India releases sovereign EKA-1 and Dhīr foundation models

India has introduced its first sovereign foundation model suites, designated EKA-1 and Dhīr, under its national artificial intelligence mission.

The releases address long-standing architectural deficiencies in how frontier systems process South Asian scripts. Frontier models trained predominantly on English corpora allocate disproportionately small vocabulary allocations to non-Latin scripts. This imbalance forces higher tokenisation ratios across Devanagari, Dravidian, and other regional scripts, inflating inference costs and compounding context window degradation for domestic users.

The EKA-1 and Dhīr model families counter this bottleneck through native Indic tokenisation and targeted pre-training across regional linguistic corpora. By expanding vocabulary support for major constitutional languages, the architectures aim to reduce byte-fallback overhead and improve generation latency across low-resource dialects. The training pipelines prioritise balanced cross-lingual representation, minimising reliance on synthetic English translations that often introduce semantic drift.

The initiative reflects a broader realignment in state-backed artificial intelligence policy across emerging economies. Rather than subsisting on proprietary application programming interfaces from United States and Chinese providers, New Delhi is subsidising native model weights and domestic compute clusters. Sovereign weight control protects domestic governance workflows from external policy shifts, export controls, and service revocations.

Compute allocation remains the central technical hurdle for the programme. Training competitive foundation models requires sustained high-bandwidth memory clusters and reliable access to cutting-edge accelerator silicon. While the national mission provides centralised access to high-performance infrastructure, domestic research consortiums must still prove their architectures can match the parameter efficiency and downstream task performance of established open-weight baselines.

Early architectural details indicate a pragmatic deployment strategy. Rather than competing directly on raw parameter scale with frontier Western systems, the developers have prioritised domain-specific fine-tuning, cross-lingual transfer, and inference efficiency for public sector and enterprise workflows across the subcontinent. How effectively these models balance cross-lingual alignment against the raw reasoning depth of larger frontier architectures will determine their long-term enterprise adoption.