SUMMARY
- Sarvam AI launches Sarvam-1, an LLM optimized for Indian languages.
- The model features 2 billion parameters and supports 10 Indic languages.
- Sarvam-1 exhibits superior performance and efficiency, making it ideal for practical applications.
Bengaluru-based Sarvam AI has unveiled Sarvam-1, a homegrown Large Language Model (LLM) designed to enhance linguistic diversity in AI applications. With approximately two billion parameters, Sarvam-1 is specifically optimized for 10 major Indic languages alongside English, aiming to foster local AI development.
According to the company’s blog, “Sarvam-1 demonstrates that careful curation of training data can yield superior performance even with a relatively modest parameter count.” The model supports languages including Bengali, Marathi, Tamil, and Telugu, and claims to surpass standard benchmarks, achieving high accuracy in both knowledge and reasoning tasks, particularly in Indic languages.
Benchmark Comparison: Sarvam-1 outperforms Gemma-2-2B and Llama-3.2-3B across various benchmarks, including MMLU, Arc-Challenge, and IndicGenBench, while achieving comparable results to Llama 3.1 8B.
Efficiency: The model boasts a 4-6X faster inference speed compared to larger models, making it suitable for practical applications, including those on edge devices.
Sarvam-1 is developed using domestic AI infrastructure powered by NVIDIA H100 Tensor Core GPUs in collaboration with partners like NVIDIA, Yotta, and AI4Bharat. The startup also offers enterprise clients solutions for speech-to-text, text-to-speech, translation, and data parsing.
