Bengaluru-based Sarvam AI has unveiled Sarvam-1, an innovative Large Language Model (LLM) focused on enhancing linguistic diversity in AI applications. Designed with approximately two billion parameters, Sarvam-1 is optimized for ten prominent Indic languages, including Bengali, Marathi, Tamil, and Telugu, as well as English, fostering advancements in localized AI technology.
In a company blog post, Sarvam AI highlighted that careful data curation allows Sarvam-1 to deliver high performance in knowledge and reasoning tasks, particularly within Indic languages, despite its relatively modest parameter count. The model’s benchmark results surpass competitors like Gemma-2-2B and Llama-3.2-3B in tests such as MMLU, Arc-Challenge, and IndicGenBench, demonstrating notable accuracy and efficacy in local contexts.
Sarvam-1’s efficiency is another standout, with a 4-6X faster inference speed compared to larger models. This optimization makes it suitable for practical applications, including those on edge devices, which are crucial for expanding AI access across diverse regions.
Developed with NVIDIA H100 Tensor Core GPUs, Sarvam-1 leverages partnerships with NVIDIA, Yotta, and AI4Bharat. Beyond language processing, Sarvam AI also offers enterprise solutions in speech-to-text, text-to-speech, translation, and data parsing, supporting versatile applications for India’s growing digital landscape.
Key Highlights:
- Sarvam AI launched Sarvam-1, a homegrown LLM supporting ten Indic languages and English.
- The model achieves superior accuracy in Indic languages and operates at 4-6X faster inference speeds.
- Partnerships with NVIDIA, Yotta, and AI4Bharat contributed to Sarvam-1’s development and efficiency.