Discuss how small, on-device language models can make AI more inclusive and secure in India. What challenges remain?

Small language models (SLMs) have fewer parameters than frontier models. They run on a phone or a local server rather than in a distant cloud. The IndiaAI Mission's Dhīr 2B/5B (ZenteiQ), released in November 2026, run on mobile phones and work offline [1]. The Economic Survey also favours decentralised, application-driven systems over capital-intensive frontier models [2]. So SLMs are a realistic route to AI that works for India, though some gaps remain.

Making AI inclusive

• Language: Dhīr supports 10 Indian languages plus English [1]. This opens AI to citizens who do not speak English.
• Connectivity: Offline use on phones reaches low-connectivity districts that cloud AI cannot serve [1].
• Affordability: Small models avoid per-query cloud fees and need less compute. This suits panchayats, small firms and schools [2].
• Open access: AIKosh hosts hundreds of models for students and researchers [3].

Making AI secure

• Privacy: Data stays on the device [1]. This fits the data-minimisation spirit of the DPDP Act, 2023.
• Sensitive use: On-premise deployment suits defence and government work, where data cannot leave the premises [1].
• Strategic autonomy: Owning the model weights removes dependence on a foreign owner's licence and updates [1].

Challenges that remain

• No independent testing: PIB says the sovereign models perform strongly on Indic benchmarks [4]. But the India AI Governance Guidelines still ask that the IndiaAI Safety Institute be resourced to build benchmarks and evaluation datasets [5].
• Capability limits: Smaller models reason less well. Errors are risky in welfare or security decisions.
• Updating offline copies: Each installed copy must be patched separately. No quick central fix is possible.
• Hardware dependence: Training GPUs are made mainly in one country. RISC-V-based Indian chips under the National Supercomputing Mission are still at an early stage [6].
• Coverage and devices: Dhīr covers 10 of the 22 scheduled languages [1], and low-end phones have limited memory.

Small, on-device models bring AI closer to every citizen and keep sensitive data at home. To build on this, the IndiaAI Safety Institute should publish an open Indic benchmark, set up secure update channels for offline copies, and link its work to the India Semiconductor Mission. Done together, these steps would deliver "Making AI Work for India" [1] and advance SDG 9 and SDG 10.