Sarvam AI launches homegrown multilingual LLM, Sarvam-1

Sarvam AI has launched Sarvam-1, a homegrown Large Language Model (LLM). With approximately two billion parameters, the model is trained on 10 major Indic languages to support linguistic diversity in building local AI applications. 

“We introduce Sarvam-1, a 2-billion parameter language model specifically optimised for Indian languages. Built from the ground up to support 10 major Indian languages, alongside English, Sarvam-1 demonstrates that careful curation of training data can yield superior performance even with a relatively modest parameter count,” read the company’s blog. 

The new model, which supports languages such as Bengali, Marathi, Tamil and Telugu, claims to outperform standard benchmarks while also achieving high accuracy on both knowledge and reasoning tasks, especially in Indic languages.

“Concretely, it easily outperforms Gemma-2-2B and Llama-3.2-3B on a variety of standard benchmarks including MMLU, Arc-Challenge, and IndicGenBench, while achieving similar numbers to Llama 3.1 8B,” the company said. 

Sarvam-1 was built from the ground up using domestic AI infrastructure powered by NVIDIA H100 Tensor Core GPUs. The LLM was developed in collaboration with several key partners, including NVIDIA, Yotta, and AI4Bharat. 

“A key advantage of Sarvam-1 is its computational efficiency: it is 4-6X faster inference speed compared to larger models while maintaining competitive performance. This combination of strong performance and superior inference efficiency makes Sarvam-1 particularly well-suited for practical applications, including on-edge devices,” said the company.

The Bengaluru-based startup provides enterprise clients with models for speech-to-text, text-to-speech, translation, and data parsing.