Sarvam’s Sovereign LLM Faces Questions After Launch
Sarvam, the AI startup chosen under the IndiaAI Mission to build India’s foundational LLM, recently launched Sarvam-M. The 24-billion parameter language model aims to power India’s AI future. But while the model sets new technical benchmarks, its reception sparked a wave of criticism and deeper debates about India’s AI ambitions.
Sarvam-M Targets Indian Language Gaps
Built on Mistral Small, Sarvam-M is a hybrid model fine-tuned heavily for Indian languages, mathematics, and programming. The company claims this model improves significantly on its French base, especially in Indic contexts. About one-third of its training data was in Indian languages. Of this, Hindi made up 28%, and nine other Indian languages like Tamil, Bengali, and Marathi each contributed 8%. Together, these account for the first language of over 70% of India’s population.
Designed for conversational AI, machine translation, and education, Sarvam-M is accessible via API and is a stepping stone toward Sarvam’s upcoming 70B multimodal AI model. Co-founder Vivek Raghavan called the launch an “important milestone” in building Sovereign AI.
Criticism Over Adoption and Approach
Despite the hype, Sarvam-M saw just 334 downloads in two days on Hugging Face. This underwhelming adoption drew fire from the startup community. Investor Deedy Das called it “embarrassing,” arguing that no one asked for an incremental 24B model. He cited Korean students’ open-source model that gained over 200,000 downloads as a contrast.
Das’s remarks ignited strong reactions. He argued Sarvam’s $41 million funding and $111 million valuation were not matched by impact. He suggested focusing on fundamental infrastructure like China’s DeepSeek instead of releasing “slightly better” models. He also questioned the need for such models when better alternatives from Google and TWO.ai exist.
Support and Backlash in Equal Measure
While Das criticised the download metrics, others defended Sarvam’s methodology and vision. Aashay Sachdeva from Sarvam AI pointed to new benchmarks achieved by Sarvam-M in Indian languages. He shared results where the model answered all JEE Advanced 2025 questions in Hindi correctly.
Others noted the importance of focusing on Indian-language AI. From farmers to legal professionals, early use cases of Sarvam-M hint at its potential utility across sectors. Critics of Das argued that building foundational models is a long-term play, not a short-term popularity contest.
A Broader AI Dilemma in India
Sarvam is not alone. BharatGen, a government-backed initiative, recently released Param 1, a 2.9B bilingual model built from scratch. Yet, it received only 12 downloads, showing that adoption remains a challenge across the board.
The larger question remains: can Indian startups deliver meaningful AI for Bharat? With 600 million smartphone users and a majority using Indic languages, the demand is clear. But compute costs, language diversity, and lack of infrastructure pose real barriers.
AI experts like Raj Dabre and Pratyush Choudhury believe the Indic language space needs homegrown models. Sarvam’s release, despite criticism, is still seen as a significant attempt. As Raj Dabre put it, “Before Sarvam-M, people complained about lack of Indic LLMs. Now they complain about the one we have.”
What’s Next for Sarvam?
The company earlier launched Bulbul, a speech model supporting 11 Indian languages with regional accents. More releases are expected in the coming weeks, as Sarvam continues its work on the ambitious 70B-parameter model.
The debate around Sarvam reflects India’s bigger AI moment. It’s not just about the number of downloads, but whether the country’s AI push can serve its diverse population. For now, Sarvam’s vision faces both optimism and scrutiny.
