Airavata
AI4Bharat's open 7B Hindi instruction-tuned research model
Overview
Airavata is a 7B Hindi instruction-following model released by IIT Madras's AI4Bharat in January 2024, LoRA fine-tuned from OpenHathi (Llama 2) on the open IndicInstruct dataset. Free gated weights on Hugging Face under the Llama 2 licence; a research model, not updated since March 2024.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 1 Oct 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
- Llama 2 Community Licence
- 7B LoRA fine-tune of OpenHathi
- Safetensors and GGUF files
- Apache-2.0 training and evaluation code
Use cases
What you can produce with Airavata
- Open 7B Hindi instruction-following model weights
- IndicInstruct dataset of Hindi and English instruction pairs
- Reproducible LoRA fine-tuning code (AI4Bharat/IndicInstruct)
- Airavata Evaluation Suite for Hindi LLM benchmarking
- GGUF file for local inference runners
ASEAN Perspective
Airavata in Southeast Asia
Airavata is an India-first model with limited direct ASEAN language coverage, as it targets Hindi and other scheduled Indian languages rather than Malay, Thai, Vietnamese, Tagalog, or Bahasa Indonesia. However, it is academically significant for ASEAN NLP researchers: its IndicInstruct methodology — clean, GPT-4-data-free instruction tuning on a 7B base — is a directly transferable blueprint for building instruction-tuned models in low-resource Southeast Asian languages. Singapore's A*STAR was a collaborating institution on the Airavata paper, indicating regional research ties. Teams at AI4Bharat and Singapore's research ecosystem have co-published, making Airavata a useful reference point even for APAC language technology builders who are not targeting Hindi.
Airavata is an honest, well-documented research release: an open Hindi instruction-tuned model, with its dataset, training code and evaluation suite all public, from AI4Bharat at IIT Madras. Its own authors say it writes natural Hindi but trails GPT-4 well behind on instruction following, is weak at maths and should not be used in production, and it has not been updated since March 2024. Use it to study or reproduce Hindi instruction tuning, or as a baseline; for a maintained Indian-language model to build products on, look at newer open-weight releases such as Sarvam AI's. Downloading needs a Hugging Face login and acceptance of the Llama 2 licence.
What we found
Airavata is a Hindi instruction-following language model released in January 2024 by AI4Bharat, the open Indian-language AI lab at IIT Madras's Nilekani Centre, which was founded in 2020. It is a seven-billion-parameter model made by LoRA fine-tuning of Sarvam AI's OpenHathi, itself built on Meta's Llama 2, using IndicInstruct: about 385,000 examples, mostly machine-translated into Hindi with IndicTrans2 and kept only if they translated back well, plus samples of real chatbot conversations and translation pairs. The authors say they used no data generated by proprietary models such as GPT-4. Weights, training code and the dataset are public, the Hugging Face repository includes a GGUF file for local runners, and the paper lists collaborators including Flipkart, Microsoft, IBM Research, NICT and Singapore's A*STAR.
The team's evaluation is candid. In a small human study of 50 prompts, each rated by a single annotator, Airavata trailed GPT-4 significantly on instruction following and content quality but wrote more natural-sounding Hindi than GPT-4 or ChatGPT; the authors call the study not robust. On the IndicXNLI inference test it improves sharply on OpenHathi, 74.7 against 48, while on IndicSentiment it slightly trails its base model. Maths is weak, at 4.02 on GSM8k. AI4Bharat describes it as a research model not recommended for production.
It has not been updated since March 2024, and AI4Bharat's releases since then have focused on translation, speech and encoder models rather than new chat models. Developers who need a maintained Indian-language model today are more likely to look at Sarvam AI's open-weight models, which cover 22 Indian languages.
Airavata is free to download from Hugging Face after logging in and agreeing to share contact details, under the Llama 2 Community Licence. There is no hosted API, paid tier or vendor support. It remains a well-documented reference for Hindi instruction tuning and for the datasets and evaluation suite released with it.
Compiled by RECATOOLS from public sources: vendor documentation and pricing, code repositories and published reporting.
Notable facts
- Airavata is named after the mythical white elephant of Hindu lore — a fitting choice for a model designed to carry the weight of Hindi AI into the future.
- The IndicInstruct dataset contains 74.7 million prompt-response pairs across 20 Indian languages, one of the largest open instruction corpora ever built for non-English languages.
- Singapore's A*STAR was a co-author on the Airavata paper — a quiet sign of ASEAN's academic investment in South Asian language AI.
- Nandan Nilekani, the architect of Aadhaar (India's 1.3-billion-person biometric ID system), has committed INR 70 crore in philanthropic grants to AI4Bharat, the lab behind Airavata.
Frequently asked questions
About this listing
This entry was compiled from publicly available data including Airavata's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with Airavata unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to Airavata directly →
Spotted something out of date? Suggest an update →
More in LLMs & Chat
