From IndicTrans2 to BharatGPT, India's homegrown AI models are setting new standards in global language tech.
India's open-source AI ecosystem has grown fast, with homegrown models across translation, speech and large language processing now competing globally. The push from IndiaAI and AIKosh has taken it even further.
Based on research across Hugging Face, GitHub, AIKosh and other repositories, this is a comprehensive list of India's most downloaded and widely used open-source AI models as of October 2025.
India's flagship translation model, IndicTrans2, supports all 22 scheduled languages and over 110 translation directions. The En-Indic 1B variant gets about 13,554 downloads per month on Hugging Face. It's the go-to model for Indian language translation tasks.
IndicBERT, a multilingual ALBERT model trained on 12 Indian languages with around 9 billion tokens, records close to 18,444 downloads per month and is used for sentiment analysis, text classification and entity recognition.
Airavata is a Hindi instruction-tuned model fine-tuned from OpenHathi. It performs well on Hindi NLP benchmarks and has around 281 downloads per month.
A multilingual speech model trained on 40 Indian languages, IndicWav2Vec represents the widest linguistic diversity among Indian automated speech recognition (ASR) models. The Hindi model alone gets about 1,997 monthly downloads.
Sarvam-1 is a two-billion-parameter language model optimised for 10 major Indic languages, including Hindi, Tamil, Bengali and Marathi. Released by Sarvam AI, the first startup to get selected under the IndiaAI Mission, the model delivers strong multilingual results across Indian contexts.
Sarvam-M is a 24 billion-parameter multilingual model built by Sarvam AI for reasoning tasks in Indic languages. Despite controversy over download manipulation, it remains a significant release in India's AI scene.
One of India's OG LLMs built on Llama, OpenHathi is a Hindi-English bilingual model based on Llama 2 with expanded tokeniser support for Indian scripts. It averages 1,733 monthly downloads and rivals GPT-3.5 on Hindi tasks.
Krutrim's first AI model built on Mistral architecture, Krutrim-1, is a seven-billion-parameter multilingual foundation model trained on two trillion tokens across Indic languages. It is a core part of Krutrim's AI stack.
BharatGPT-3B-Indic is a sovereign transformer-based language model optimised for Indic languages. It crossed 2,000 downloads within days of release. Quantised GPT-generated unified format versions receive around 129 monthly downloads.
Param-1 is one of the first Indian LLMs built from scratch by the BharatGen consortium. It is a 2.9-billion-parameter bilingual foundation model trained on 7.5 trillion tokens in English and Hindi. The instruct variant records about 4,369 monthly downloads, making it one of the most active Indian LLMs.
Project Indus, an open-source Hindi LLM supporting 37 dialects, is focused on enterprise applications and local data understanding. The LLM is built from scratch, and Tech Mahindra has also released a second version of the model.
MuRIL is a BERT-based model trained on 17 Indian languages and their transliterated counterparts. It remains one of the most widely used Indian language NLP models worldwide.
Vakyansh ASR is an automatic speech recognition model trained for multiple Indian languages. The Hindi model gets 2,838 monthly downloads, while other languages cover Tamil, Kannada, Telugu and Bengali.
IndicTTS is an open-source text-to-speech model for 13 languages, including Hindi, Tamil, Malayalam and Marathi. It is widely used in Indian digital-voice applications.
Indic Parler-TTS is a multilingual text-to-speech model supporting 21 languages and trained on 1,806 hours of data from 69 unique voices. It leads Indian TTS downloads with 12,295 per month.
The Paramanu family of models consists of lightweight models (13 M-367 M parameters). They are trained for 10 Indian languages, optimised for minimal compute.
Paramanu-Ayn is India's first legal AI model, trained on Supreme Court judgments, the Constitution and the Indian Penal Code documents. That apart, Paramanu-Ganita is a math-focused AI model designed for Indian education and tutoring applications.
Chirtrarth is a vision-language model combining Krutrim-1 with a SigLIP visual encoder. It supports 10 Indic languages along with English for multimodal reasoning.
e-vikrAI is a vision-language AI for Indian e-commerce that automates product cataloging using image and text inputs with local cultural accuracy. The model simplifies cataloguing for sellers, eliminating the need for manual input.