Sarvam-Translate
Document-level translation model for 22 Indian languages.
Sarvam-Translate is a translation model created by Sarvam AI and AI4Bharat for 22 official Indian languages. Built on the Gemma3-4B-IT base, it processes full documents and complex formats rather than single sentences, aiming to improve low-resource language performance.
- Translates across 22 official Indian languages and English.
- Handles full documents with an 8192 token context window.
- Built on the four billion parameter Gemma3-4B-IT base model.
- Processes markdown, optical character recognition errors, and equations.
- Requires English pivoting for translation between two Indian languages.
What it is
Sarvam-Translate is a language model designed for translating text across 22 official Indian languages and English. Developed jointly by Sarvam AI alongside AI4Bharat, it focuses on document-level translation rather than processing isolated sentences. The model is built on the Gemma3-4B-IT architecture, giving it four billion parameters. It aims to close the performance gap for Indian languages in large language models. The system handles diverse content types, including web pages, textbooks, and scientific articles. It also processes varied formats like markdown, text with optical character recognition errors, and documents containing mathematical or chemical equations. For code files, it translates only the comments.
How it came about
The model was created through a partnership between Sarvam AI and AI4Bharat. The developers built it to address the specific needs of Indian languages, which historically receive less attention in large language model development compared to high-resource languages. By leveraging the Gemma3-4B-IT base, the team fine-tuned the system for comprehensive translation tasks. The open-sourced checkpoint was released on the Hugging Face platform, allowing researchers and developers to download and test the weights directly. The team also made the model available through proprietary application programming interfaces for commercial or integrated use.
Inside the model
The architecture relies on a decoder-only transformer design inherited from its base model, Gemma3-4B-IT. It operates with a maximum context window of 8192 tokens, allowing it to read and translate long documents in a single pass. The training process emphasises context awareness and inclusivity, ensuring the model respects different styles, formats, and appropriate gender attribution. The model can be deployed locally using vLLM or the Hugging Face Transformers library, and it supports bfloat16 precision for efficient processing.
What it can do
The model excels at translating entire documents, maintaining context across long passages. It supports 23 languages in total, covering the 22 official Indian languages plus English. It handles complex inputs like digitised text with recognition errors and embedded equations. The developers engineered it to provide nuanced and accurate translations, ensuring inclusivity and appropriate gender attribution across different contexts and formal or informal styles.
How to use it
The open-sourced checkpoint is available for download on the Hugging Face platform under an open licence. Developers can run the model locally using the Transformers library or deploy it as a server with vLLM. Additionally, Sarvam AI provides access through its own application programming interfaces, offering a Python client for integration into external applications.
Safety and testing
The developers engineered the model to be context-aware and inclusive. A key focus of its design is ensuring appropriate gender attribution during translation. The system is built to respect different contexts, formats, and styles, avoiding biased or inappropriate outputs when handling formal or informal text.
Reception and debate
The provided material does not detail public reception, critical reviews, or market adoption. The model is presented as a solution to the historical lag in large language model performance for Indian languages, aiming to provide high-quality, contextually aware translations for these regions.
Why it matters
By focusing on document-level translation and complex formats, the model aims to improve the availability of high-quality translation tools for Indian languages. Its open-sourced nature allows the research community to build upon its weights, potentially advancing the field of low-resource language processing and machine translation in the region.
Questions people ask about Sarvam-Translate
What languages does Sarvam-Translate support?
The model supports 23 languages in total. This includes English and the 22 official Indian languages, such as Hindi, Tamil, Telugu, Bengali, Marathi, Gujarati, and Sanskrit. To translate between two Indian languages, the system must first translate the text into English and then into the final target language.
How long can the input documents be?
The model has a maximum context length of 8192 tokens. This allows it to process and translate entire documents, web pages, speeches, and scientific articles in a single pass, rather than being restricted to short, isolated sentences.
Can it translate code files?
Yes, the model can process code files, but it is designed to translate only the comments within the code. It leaves the actual programming syntax and logic untouched while providing translations for the human-readable explanatory text.
How is the model made available to users?
Users can access the open-sourced checkpoint directly from the Hugging Face platform to run it locally. Alternatively, Sarvam AI offers the model through its own application programming interfaces, providing a Python client for easier integration into software applications.
Good reading and sources
There is no English Wikipedia article on Sarvam-Translate yet, so the facts on this page come from the sources listed above: huggingface.co/sarvamai/sarvam-translate. 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 2026-11-15.