Sarvam-1
A 2-billion parameter language model for Indian languages.
Sarvam AI released Sarvam-1 in October 2024. This 2-billion parameter model focuses on ten Indian languages and English. It uses a custom tokenizer for better efficiency and was trained on high-quality Indic data.
- Optimized for 10 Indian languages and English.
- Uses 2 billion parameters with 8,192 token context.
- Trained on 2 trillion high-quality Indic tokens.
- Achieves 4-6x faster inference than larger models.
- Released under a non-commercial license.
What it is
Sarvam-1 is a language model with 2 billion parameters. Sarvam AI built it to serve speakers of Indian languages. The system supports ten major Indic tongues alongside English. It acts as a base model for text completion rather than a chat assistant. Users must fine-tune it for specific tasks. The company claims it matches larger models in its target languages while running faster. It aims to close the technological gap for billions of Indic speakers.
How it came about
Sarvam AI announced the model in October 2024. The team noted that most large models focus heavily on English. They sought to address poor token efficiency and low data quality for Indic scripts. Previous efforts often relied on web-crawled data with limited depth. This project started from scratch to support Indian languages properly. The developers partnered with AI4Bharat for expertise in these technologies.
Inside the model
The architecture uses a decoder-only transformer design. It features 16 attention heads and SwiGLU activation functions. Positional embeddings use rotary encoding with a theta of 10,000. The context window extends to 8,192 tokens. Training employed grouped-query attention and bfloat16 mixed precision. NVIDIA NeMo Framework facilitated the process on Yotta’s Shakti cluster. The run used 1,024 HGX H100 GPUs over five days. A custom tokenizer with 68,096 tokens improves fertility rates significantly.
What it can do
The model handles Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Oriya, Punjabi, Tamil, and Telugu. It achieves fertility rates between 1.4 and 2.1 for these scripts. This is two to four times more efficient than other multilingual systems. Benchmarks show it outperforms Gemma-2-2B and Llama-3.2-3B on translated tasks. It remains competitive with Llama-3.1-8B despite being much smaller. Inference speeds are four to six times faster than those larger counterparts.
How to use it
Weights are available on the Hugging Face Hub. The licence permits non-commercial use only. Developers can download the model and tokenizer for local deployment. It requires fine-tuning for downstream applications. The company suggests it suits edge device deployment due to its size. Direct chat or instruction following is not supported without further training.
Safety and testing
The training corpus reduced coverage of sensitive topics by half compared to web data. The dataset prioritised scientific and technical content. This approach aimed to create a responsible foundation for Indic AI. The company highlighted lower repetition rates and improved coherence in the data. No specific red-teaming results or safety incidents are detailed in the provided material.
Reception and debate
The release positioned the model as the first Indian language LLM built from the ground up. Claims focus on superior performance relative to its parameter count. Comparisons highlight efficiency gains over established multilingual models. The work addresses long-standing issues with tokenisation for Indic scripts.
Why it matters
Sarvam-1 demonstrates that careful data curation can yield strong results with modest scale. It challenges the dominance of English-centric development in large language models. The custom tokenizer offers a template for better efficiency in other scripts. The accompanying Sarvam-2T corpus provides a high-quality resource for future research. It shows that smaller models can compete with larger ones in specific domains.
Questions people ask about Sarvam-1
Which languages does Sarvam-1 support?
It supports English and ten Indian languages: Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Oriya, Punjabi, Tamil, and Telugu. The model was built specifically to handle these scripts efficiently.
Can I use Sarvam-1 for commercial projects?
No. The model is released under the Sarvam non-commercial license. You may download the weights for research or personal use, but commercial deployment requires a different agreement or is not permitted.
Is Sarvam-1 a chatbot?
No, it is a text-completion base model. It cannot follow instructions or chat directly. Developers must fine-tune it on specific datasets to create instruction-following or conversational applications.
Good reading and sources
There is no English Wikipedia article on Sarvam-1 yet, so the facts on this page come from the sources listed above: huggingface.co/sarvamai/sarvam-1, sarvam.ai/blogs/sarvam-1. 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.