Two AI, a new startup in the artificial intelligence (AI) space, has launched SUTRA, an innovative language model capable of proficiency in over 30 languages. It includes many South Asian languages such as Gujarati, Marathi, Tamil, and Telugu, aiming to address the unique linguistic challenges and opportunities in South Asia.
Constructed by using two mixture-of-experts transformers – a concept model and an encoder-decoder for translation – SUTRA has been developed to predict the next token. It primarily uses publicly accessible datasets in languages with abundant data, such as English, while its translation model is trained on 100 million human and machine translated conversations across multiple languages. This method allows the model to map concepts to similar embeddings in all the languages it supports.
Users can translate an input text into an initial embedding using the translation model’s encoder. The concept model then processes this text and feeds it into the translation model’s decoder to produce the final output. This effective integration ensures that SUTRA can handle a wide range of languages, marking it a robust solution for multilingual communication.
SUTRA is offered in Pro, Light, and Online versions. Both SUTRA-Pro and SUTRA-Online provide high performance and internet connectivity at a cost of $1 per 1 million tokens, while SUTRA-Light offers a low-latency solution at $0.75 per 1 million tokens. The model’s competitive pricing makes it a popular choice for users and businesses in cost-sensitive areas.
Performance-wise, SUTRA has shown impressive results, especially in the South Asian context. On the multilingual MMLU benchmark, SUTRA outperformed GPT-4 in Gujarati, Marathi, Tamil, and Telugu. Its tokenizer also displays superior efficiency, producing fewer tokens than GPT-3.5 and GPT-4, particularly in non-Latin languages such as Hindi and Korean. However, SUTRA’s evaluation only currently covers 11 out of its 33 languages, hinting at the potential for further validation and improvement.
Two AI’s specific focus on non-English-speaking markets such as India, Japan, South Korea, and the Middle East is backed by significant $20 million seed funding from Jio and Naver. The company is well-placed to disrupt the AI market by offering a competitively priced model that excels in local languages.
While there is room for refinement, SUTRA’s targeted performance, efficiency, affordability, and high-quality multilingual support position it as a strong competitor in the multilingual AI landscape. With the capacity to bridge the gap for users in underrepresented areas, SUTRA could play a critical role in the global AI landscape.
