Sarvam AI (Axonwise Pvt Ltd), India’s leading AI startup, recently launched its newest LLM (large language model) named Sarvam-M has caused a debate about its approach and the problems with India’s AI (artificial intelligence).

Sarvam AI, an early company selected for the IndiaAI Mission, launched Sarvam-M. Based on Mistral Small, this 24-billion parameter open-source model is a key development for Indic AI. It supports 10 Indian languages, including Hindi, Bengali, Gujarati, Kannada, and Malayalam.

A limited 334 downloads in two days on Hugging Face led to considerable backlash for the project. Deedy Das, an investor at Menlo Ventures, openly criticised this model as “embarrassing”, and his comments, saying this small improvement wasn’t necessary, have sparked a heated debate among Indian AI experts.

Das highlighted the contrast with an open-source model created by two Korean college students, which achieved approximately 200,000 downloads.

Even though Sarvam promised more models later, their first one, Sarvam-M (which used a French AI model), faced criticism. Sarvam AI isn’t alone in getting few downloads. The government’s Param-1 model only had 12 downloads last week.

Das’s criticism isn’t just about downloads. He thinks Sarvam’s work shows they have the wrong goals. He said, “No one is asking for a slightly better 24B Indic model. “If you want to train models, there should be a very good reason for it.”

He added that Google and TWO.ai have models that are both cheaper and perform better in all these languages. He said, “I have nothing against Sarvam, but I just don’t think that at this moment their contributions are remotely commensurate to their funding”.

Sarvam has secured $41 million from investors like Lightspeed India Partners and Khosla Ventures, reaching a $111 million valuation by March 2025. Despite this, X users noted the model’s utility but also its need for improvement. Das believed Sarvam, despite their good Bulbul TTS model, needed to completely change how their systems work, pointing to DeepSeek in China as a model to follow.

Sarvam-M, despite low initial downloads, shows strong performance in Indian languages, outperforming some established models. While it gained criticism for its limited downloads, Sarvam AI emphasises its special role in improving AI for Indian languages and solving problems unique to India, like helping farmers. 

The debate highlights the challenge and importance of building AI for India. We need AI that understands India’s many languages and helps its large population who speak those languages, instead of just aiming for global fame.

Source: https://analyticsindiamag.com/ai-features/sarvam-ais-backlash-exposes-the-sad-state-of-indian-ai/