Introduction

In the world of big-league tech, where giant global players usually lead the AI race, India is making some exciting moves of its own. A whole new world of Indian-made Large Language Models (LLMs) and AI tools is starting to shine, each with its own special flair. We’re here to put these local heroes under the spotlight, showing off their cool features and groundbreaking progress. Ready for an adventure into the diverse and dynamic world of India’s own AI creations? Let’s jump in and discover what makes these Indian LLMs and AI tools not just smart, but truly remarkable. Let us now look on top 10 LLMs built in India.

 Navarasa 2.0

Telugu LLM Labs presents Navarasa 2.0, an advanced iteration of the Gemma series language models. This model, a 7B/2B instruction-tuned configuration, supports an extensive suite of 15 Indian languages along with English, building upon its predecessor that was initially fine-tuned for 9 Indian languages.

Navarasa 2.0 is designed to be versatile, suitable for various applications including content generation, translation, customer support, and educational resources, particularly in local languages. Its capability to function across multiple Indian languages substantially increases its utility for businesses and developers targeting India’s linguistically diverse population. Additionally, this broad language support plays a crucial role in enhancing digital inclusivity, allowing more individuals to access technology in their native languages.

Key Features

• Base Models: Utilizes the Gemma 7B/2B models as the foundation for fine-tuning.
• Expanded Language Portfolio: Includes additional languages such as Marathi, Urdu, Konkani, Assamese, Nepali, and Sindhi, bringing the total to 16 languages including English.
• Data Enrichment: Employs a translated version of the alpaca-cleaned-filtered dataset, now extended to cover six more Indian languages, enhancing the training breadth and depth.
• Enhanced Generative Capabilities: The model has been specifically enhanced to bolster its generative abilities, promoting more effective and context-aware text generation across multiple languages.

Click here to explore Navarasa 2.0.

Languages Supported by Navarasa 2.0

Hindi, Telugu, Tamil, Kannada, Malayalam, Marathi, Gujarati, Bengali, Punjabi, Odia, Urdu, Konkani, Assamese, Nepali, Sindhi, and English.

With Navarasa 2.0, Telugu LLM Labs underscores its commitment to reducing linguistic barriers and fostering a more inclusive digital environment in India. This model exemplifies the potential of AI to cater to and enrich the multilingual fabric of the Indian subcontinent.

Dhenu 1.0

Expanding its portfolio of transformative AI solutions for agriculture, KissanAI proudly introduces Dhenu, a series of Language Learning Models (LLM) directly inspired by the mythological Kaamdhenu—the wish-fulfilling cow from Hindu mythology. Dhenu represents the epitome of marrying tradition with cutting-edge technology, designed specifically to serve the agricultural sector with precision and innovation.

Dhenu-vision-lora-v0.1, a part of this series, is an open-source agricultural disease detection model that has been fine-tuned using the Qwen-VL-chat model. This model is crafted to assist farmers in identifying diseases in three major crops—rice, maize, and wheat—through a conversational interface, integrating advanced Low Rank Adaptation techniques for cost-effective fine-tuning on specialized agricultural datasets.

Key Features

• Model Base: Utilizes the “Qwen/Qwen-VL-Chat” as its base LLM, with enhancements through the LoRA methodology.
• Specific Focus: Tailored to enhance disease detection capabilities in agriculture, significantly outperforming the base model with a 2X improvement.
• Crop Diseases Addressed: The model can identify a variety of diseases across rice, maize, and wheat, such as Leaf Blight, Leaf Spot, and Wheat Loose Smut, among others.
• Training and Dataset: Trained in March 2024, using a synthetic dataset of approximately 9,000 images highlighting common crop diseases.
• Evaluation: Tested on a collection of 500 images, Dhenu-vision-lora-v0.1 achieved a 36.13% accuracy rate, demonstrating substantial advancements over the base model.

Click here to access Dhenu 1.0

Odia Llama

The OdiaGenAI team has released a fine-tuned Llama2 model dedicated to the Odia language, addressing the linguistic nuances and cultural specifics of Odisha. This model enhances the digital presence of the Odia language, which has historically been underrepresented in AI applications.

Key Features

• Rich Training Dataset: Incorporates a diverse range of domain-specific knowledge, covering topics from local cuisine to historical sites.
• Advanced Fine-Tuning: Utilizes Low-Rank Adaptation (LoRA) methods tailored for Odia, improving the model’s performance on native content.
• Cultural Relevance: Trained to respect and reflect the cultural heritage of Odisha, ensuring that the generated text resonates with local users.
• Accessible and Open-Source: Available for research and non-commercial use, promoting further academic and practical exploration.

Explore the full discussion at OdiaGenAI.

Kannada Llama

Tailored for the Kannada-speaking community, Kannada Llama enhances AI’s linguistic capabilities in handling the Kannada language. This model is meticulously engineered to support diverse applications, from conversational AI to text analysis.

Key Features

• Extensive Training: Pre-trained on over 600 million Kannada tokens to capture nuances of the language.
• Advanced Techniques: Utilizes Low-Rank Adaptation (LoRA) for efficient training and fine-tuning.
• Optimized Datasets: Fine-tuned on specialized datasets to improve conversational capabilities and text comprehension.
• Open-Source Contribution: Facilitates wider access and collaboration within the tech community, promoting further research and development in Indic language AI.

Explore more details on Kannada Llama at Tensoic Blog.

OpenHathi

OpenHathi, with its name meaning “elephant” in Hindi, is not just a large language model, but a symbol of the growing power of Indian languages in the AI landscape. This 7B parameter model, developed by Sarvam AI, marks the first release in the OpenHathi series, designed to empower diverse applications in the Indian market. As the first publicly available Hindi Large Language Model (LLM), OpenHathi represents a pivotal moment in India’s AI evolution. 

Key Features

• Bilingual Training: OpenHathi leverages not just Hindi but also English and Hinglish data during training, enhancing its comprehension and generation capabilities across both languages.
• Custom Tokenization: A unique sentence-piece tokenizer with a 16K Hindi vocabulary merges with the Llama2 tokenizer to significantly reduce tokenization overhead for Hindi text.
• Phased Training: The model undergoes a three-phase training process:
  • Phase 1: Bilingual text translation using low-rank adapters, fostering cross-lingual understanding.
  • Phase 2: Bilingual next-token prediction with low-rank adapters, enabling context-aware language generation.
  • Phase 3: Supervised fine-tuning on internal datasets for specific tasks, tailoring the model’s ability to handle diverse applications.
• Open-source Accessibility: The OpenHathi base model after phase 2 is publicly available via HuggingFace, allowing developers and researchers to fine-tune it for their specific needs and tasks.
• Cross-lingual Potential: OpenHathi’s bilingual training opens doors for potential applications in cross-lingual translation, information retrieval, and other tasks that require seamless interaction between Hindi and English.

Click here to explore OpenHathi.

Tamil-LLAMA

Tamil-LLAMA is a large language model specifically designed for the Tamil language. Developed by Abhinand Balachandran, it builds upon the foundation of the LLaMA model but significantly enhances its capabilities in handling Tamil text.

Key Features

• Enhanced vocabulary: The model’s vocabulary expands upon the original 32,000 tokens by incorporating an additional 16,000 Tamil-specific tokens, enabling more nuanced and accurate processing of Tamil language.
• Efficient training: Leveraging the LoRA methodology, Tamil-LLAMA achieves optimal training efficiency while maintaining model robustness.
• Multiple variations: Four variations are available: Tamil LLaMA 7B, 13B, 7B Instruct, and 14B Instruct. Each variation offers different parameter sizes and fine-tuning approaches, catering to diverse needs and computational resources.
• Fine-tuning with focused datasets: To further refine its Tamil comprehension and generation abilities, the model undergoes additional training with a Tamil-translated version of the Alpaca dataset and a subset of the OpenOrca dataset, specifically chosen for their relevance to Tamil language tasks.
• Open-source availability: The code, models, and datasets are all publicly available, fostering further research and development in Tamil language processing.

Overall, Tamil-LLAMA represents a significant leap forward in the field of Tamil language AI. Its combination of enhanced vocabulary, efficient training methods, focused fine-tuning, and open-source accessibility makes it a valuable tool for researchers, developers, and anyone interested in leveraging the power of AI for Tamil language applications.

Click here to explore this LLM built in India.

Krutrim

 Krutrim AI is a generative AI assistant that converses in 10+ languages, including Hindi, English, Tamil, Telugu, Malayalam, Bengali, Marathi, Kannada, Gujarati, etc., making it India’s own AI by an artificial intelligence startup. It was founded by Bhavish Aggarwal, who is also the founder of Ola cabs. Krutrim AI has been natively created to ensure a creative AI tool designed for over 1.4 billion Indians to provide 100% contextually relevant responses. The company aims to revolutionize how Indians interact with technology, breaking down the linguistic and cultural barriers that often hinder AI adoption. Krutrim AI is currently in public beta and is poised to transform the Indian customer service landscape with its AI-powered chatbots.

Click here to explore this LLM built in India.

Project Indus

Tech Mahindra has unveiled a really cool project called Project Indus, which is all about making computers understand Hindi and its many dialects. It is at the forefront of a groundbreaking initiative in language technology, developing a pure Hindi Large Language Model (LLM) powered by AI. This model is notable for its substantial scale, encompassing 539 million parameters and a vast collection of 10 billion tokens from Hindi and its dialects. The project’s ambitious goal is to build an Open Source LLM, aiming to revolutionize language technology and meet the needs of a quarter of the world’s population. This endeavor will create extensive language repositories, promising significant benefits for sectors like rural finance, retail, and logistics, thereby contributing to growth across India.

The initial phase of Project Indus focuses on Hindi and its 37 dialects, laying a solid foundation for future expansion. Over time, the project will incorporate additional languages and dialects, broadening its scope and impact. This initiative by Tech Mahindra is more than just a technological advancement. It’s a step towards bridging language barriers and fostering inclusivity on a global scale.

Click here to explore this LLM built in India.

Bhashini

Bhashini is a significant initiative by the Government of India designed to bridge the digital divide by democratizing access to digital services across various Indian languages. This national public digital platform aims to develop services and products by leveraging artificial intelligence and other emerging technologies. Bhashini’s efforts focus not just on developing Large Language Models (LLMs), but also on creating a comprehensive ecosystem that supports language technology through various projects.

Bhashini encompasses a diverse landscape of language technology projects, with LLM development as a crucial element. This holistic approach extends beyond individual languages, seeking to create bridgepoints between technology and India’s rich linguistic heritage. Bhashini envisions digital inclusivity as a lived reality for all citizens by dismantling language barriers.

One of the key components of Bhashini is the Universal Language Contribution API, which is an open-source platform used to collect, curate, and discover datasets in Indian languages. Bhashini’s platform enhances language tech, supporting speech recognition, text-to-speech, machine translation, advancing Indian language processing.

While still in its beta phase, the Bhashini app marks a significant milestone in the program’s journey. Available for download on both Apple Store and Google Play Store, the app offers a glimpse into the transformative potential of Bhashini. As the program grows, it will impact education, healthcare, governance, and economic development domains.

Click here to explore this made in India LLM.

BharatGPT

BharatGPT, by CoRover.ai, is a transformative Generative AI platform tailored for the Indian market. It supports over 14 languages across various modalities. Fully aligned with the Indian government’s initiative, BharatGPT ensures data sovereignty and security by keeping all data within the country. BharatGPT, versatile and integrated with ERP/CRM systems, supports multiple languages and formats, featuring an inbuilt payment gateway for real-time transactions.

BharatGPT’s multi-layered query processing reduces computational load, enhancing efficiency and scalability for diverse organizational requirements. BharatGPT is essential across sectors, utilized by major organizations like IRCTC and LIC for varied functions.

BharatGPT offers customizable experiences, including adding custom knowledge bases, appealing to enterprises seeking tailored AI solutions.

Click here to access this LLM

Conclusion

India is making big strides in artificial intelligence, particularly with its own Large Language Models and AI tools. We’ve looked at a range of exciting projects—from Navarasa 2.0, which supports many Indian languages, to Dhenu, which helps farmers detect crop diseases, and Odia Llama, which focuses on the Odia language. These projects show India’s dedication to using AI to help different regions and people.

We’ve also seen innovative projects like OpenHathi and Tamil-LLAMA that are pushing the boundaries of what AI can do in India. On top of these, ambitious initiatives like Project Indus, the Bhashini program are making technology accessible to people across India. As India continues to grow in AI, we’d love to hear about more projects.

If you’ve created any homegrown LLM or know any that deserves to be in the above list. Let me know in the comments section. Let’s talk about the exciting world of AI in India!

Himanshi Singh 19 Apr 2024

I am a data lover and I love to extract and understand the hidden patterns in the data. I want to learn and grow in the field of Machine Learning and Data Science.