OpenHathi
A 7 billion parameter language model for Hindi and English.
Sarvam AI released OpenHathi, a 7 billion parameter base model built on the Llama 2 architecture. It is trained on Hindi, English, and Hinglish data to support Indian language applications. The company advises users to fine-tune the model for specific tasks rather than using it directly.
- Built by Sarvam AI for Indian languages.
- Contains 7 billion parameters.
- Trained on Hindi, English and Hinglish.
- Based on the Llama 2 architecture.
- Requires fine-tuning for specific tasks.
What it is
OpenHathi is a large language model developed by Sarvam AI. It contains 7 billion parameters and serves as the first entry in the OpenHathi series. The model is designed to process and generate text in Hindi, English, and Hinglish, a mix of Hindi and English. It is built upon the Llama 2 architecture. As a base model, it is intended for further adaptation by developers rather than direct deployment.
How it came about
Sarvam AI introduced this model as the initial release in its OpenHathi series. The company published details regarding the training procedure and evaluation metrics on its blog. The repository hosting the model weights identifies this specific version as OpenHathi-7B-Hi-v0.1-Base. No specific release date is provided in the available documentation.
Inside the model
The model uses a decoder-only transformer structure derived from Llama 2. It was trained on a dataset comprising Hindi, English, and Hinglish text. This multilingual approach aims to capture the linguistic nuances of Indian users who code-switch between languages. The architecture supports standard causal language modeling tasks. Developers access the model using tools compatible with the Llama tokenizer and framework.
What it can do
OpenHathi is capable of processing inputs in three linguistic varieties: Hindi, English, and Hinglish. It functions as a foundational model, meaning it possesses general language understanding but lacks specific task alignment. The creators state that it requires fine-tuning to perform well on particular applications. It is not optimized for zero-shot use cases without additional training.
How to use it
The model weights are available for download via Hugging Face. Sarvam AI provides the base model under an open access scheme. Users can integrate it into their workflows using standard transformer libraries. The company explicitly recommends that developers fine-tune the model on their own datasets before deploying it for production tasks.
Questions people ask about OpenHathi
What languages does OpenHathi support?
The model is trained on Hindi, English, and Hinglish. This combination allows it to handle code-switching common in Indian digital communication. It is designed to serve users who operate in these three linguistic contexts.
Can I use OpenHathi directly for chat?
No, this is a base model. The developers recommend fine-tuning it on specific tasks before use. It is not aligned for direct conversational interaction without further training or adjustment by the user.
What architecture is OpenHathi based on?
It is based on the Llama 2 architecture. This means it uses a decoder-only transformer design. Developers familiar with Llama 2 can use similar tools and tokenizers to work with this model.
Good reading and sources
There is no English Wikipedia article on OpenHathi yet, so the facts on this page come from the sources listed above: huggingface.co/sarvamai/OpenHathi-7B-Hi-v0.1-Base. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.