India’s LLM landscape: models, research, and infrastructure

Search results often collapse model origin, language support, infrastructure, and company location into one phrase. They are separate attributes. A useful map names which layer is being discussed before comparing projects.

The Union Cabinet approved the IndiaAI Mission in March 2024. Its official components include compute capacity, an innovation centre, datasets, application development, future-skills work, startup financing, and safe-and-trusted AI. Those components matter together: model weights alone do not create a research or deployment ecosystem.

In February 2026, the Press Information Bureau reported that twelve teams had been selected to build India’s foundation models under the mission. That is evidence of a national program and selected development efforts—not proof that every selected model is released, independently benchmarked, or frontier-capable. Project-level claims should still be checked against each team’s model card, access terms, and evaluation record.

AI4Bharat at IIT Madras maintains research, datasets, models, tools, and publications for Indian languages. Its work is useful evidence for the ecosystem because it exposes language-specific resources and evaluation methods, not just a general multilingual claim. The lab’s Indic LLM Arena also documents a human-preference evaluation direction for multiple Indian languages.

BharatGen publishes a research-paper index covering its multilingual and multimodal program. These public records make it possible to follow methods and outputs over time. A directory entry here is not an endorsement of every result; readers should inspect the linked paper, release, or model card for the exact claim.

Indian companies and research programs publish models, APIs, and technical documentation at different levels of openness. Sarvam AI, for example, maintains documentation for its Sarvam-105B offering. That page can support claims about the product it documents; independent comparisons still need fixed versions, shared tasks, identical settings, and language-qualified human review.

“Open,” “open-weight,” and “open-source” also need separate checks. Weight access does not automatically disclose the training data, training code, or rights needed to reproduce and modify the system. The open-weight versus open-source guide explains the distinction.

BrokenGPT is an India-built AI chat and API platform. It exposes a stable public model alias, routes requests to configured third-party or open-weight model infrastructure, meters chat and API usage, and publishes the service boundary. That is a useful product claim; it is not a claim that BrokenGPT trained the current foundation model.

This wording also avoids confusing “frontier” with “unrestricted.” A model can be more willing to answer without being at the leading edge of broad capability, and a frontier model can have strict product policies. See the frontier LLM guide for the independent axes.

BrokenGPT’s planned original work focuses on questions a broad English leaderboard cannot answer:

• How do refusal and answer quality change between English, native script, transliteration, and code-switched prompts?
• Which safety or refusal categories produce false positives for culturally local, political, historical, or adult-health questions?
• Do lower-refusal adaptations preserve factuality, code quality, calibration, and long-context retrieval?
• How do serving precision, quantization, latency, and cost change the result for India-facing deployment?
• Which languages, dialects, domains, and user groups remain under-tested?

These are planned research questions, not completed findings. Releases will be labelled only after prompt sets, model versions, sampling settings, scoring methods, costs, and limitations are ready for inspection.