Bengaluru Startup Launches Marathi Language AI Model Amid Growing Competition

Bengaluru startup Smallstep.ai launches 'Misal', a Marathi-focused AI language model, aiming to make AI more accessible for non-English speakers in India.

author-image
Rafia Tasleem
Updated On
New Update
Bengaluru Startup Launches Marathi Language AI Model Amid Growing Competition

Bengaluru Startup Launches Marathi Language AI Model Amid Growing Competition

Bengaluru-based startup Smallstep.ai has launched 'Misal', a large language model (LLM) tailored specifically for Marathi speakers. Inspired by the popular Marathi dish 'Misal Pav', the startup intends to cater to the growing demand for AI capabilities in local languages across India.

Misal LLM is built on top of Meta's Llama2 model and comes in four versions designed to enhance Marathi language processing. Founder Sagar Sarkale recognized the gap in AI models for Marathi speakers and decided to create Misal to make AI more accessible and useful for people who speak languages other than English.

Smallstep.ai used a custom SentencePiece tokenizer to improve performance for Marathi text compared to the Llama2 model, which is primarily trained on English data. The company claims that the Misal-7B version outperformed ChatGPT 3.5 in reading comprehension tasks, although it lagged in other areas.

Why this matters: The launch of Misal marks a significant step in the development of AI capabilities for local Indian languages. As competition heats up in this space, models like Misal have the potential to make AI more inclusive and beneficial for a wider population of non-English speakers in India.

Smallstep.ai acknowledges the need for better evaluation methods to benchmark the model and is working on improvements to address issues like hallucinations and biases. The company is dedicated to continuously improving Misal, with future iterations promising enhanced techniques and safety measures to become a staple tool for Marathi speakers navigating the digital realm.

Sarkale emphasized the importance of Misal, stating, "We want to bridge the gap in AI models catering to Marathi speakers. Misal is designed to make AI more accessible and useful for people who speak languages other than English." As local language AI models gain traction, startups like Smallstep.ai are at the forefront of efforts to democratize AI technology and make it more inclusive for diverse linguistic communities in India.

Key Takeaways

  • Smallstep.ai launched 'Misal', an LLM for Marathi speakers.
  • Misal is built on Meta's Llama2 model and outperformed ChatGPT 3.5 in reading tasks.
  • Misal aims to make AI more accessible for non-English speakers in India.
  • Smallstep.ai is working to improve Misal's evaluation and address issues like hallucinations.
  • Local language AI models like Misal can democratize AI technology in India.